{ "updated": "2026-10-01", "methodology": "Provider documentation reviewed on the dates shown. This is a curated starting catalog, not an exhaustive ranking. No hands-on model tests were performed for these profiles. Unknown fields are left unknown; a review date is not a release date.", "models": [ { "slug": "gpt-6-astra", "name": "GPT-6 Astra", "provider": "OpenAI", "category": "Language model", "reviewed": "2026-09-27", "summary": "A hosted model to evaluate for reasoning and code workflows, with text output and documented image-input support.", "source": "https://developers.openai.com/api/docs/models/gpt-6-astra", "facts": [ [ "API identifier", "gpt-6-astra" ], [ "Inputs / outputs", "Text and image inputs; text output. Native audio and video are not listed as supported." ], [ "Context / maximum output", "1,050,000 / 128,000 tokens" ], [ "Standard token prices", "USD $10 input and $50 output per million tokens. Cached input is $1 per million." ], [ "Long-context caveat", "Above 272,000 input tokens, the documentation applies 2× input/cache and 1.5× output rates to the full request." ], [ "Access", "Hosted API. Downloadable weights are not documented on the reviewed model page." ] ], "decision": "Image input is not the same capability as native image generation. For a visual workflow, assess whether you need a model to explain an image or an endpoint that creates pixels. Tool-generated images also need their own model and cost recorded. For text and code evaluation, record reasoning settings alongside the model name so that a rerun can be interpreted.", "checklist": [ "Run a fixed set of tasks with recorded inputs, reasoning effort and acceptance criteria.", "Track input, cached input and output separately, including the long-context threshold.", "Compare saved text outputs in the editor; do not score an image-generation tool as if Astra itself produced the pixels." ], "limitations": "This profile does not establish latency, accuracy or superiority over another model. Tool fees, fast modes and volume discounts are outside the simple standard-price fields. Confirm billing against your selected API mode before estimating a workload." }, { "slug": "claude-opus-5-5", "name": "Claude Opus 5.5", "provider": "Anthropic", "category": "Language model", "reviewed": "2026-09-28", "summary": "A hosted model with an official September 22 announcement and separate standard, cache and fast-mode pricing.", "source": "https://www.anthropic.com/claude-opus-5-5", "facts": [ [ "Announcement date", "September 22, 2026" ], [ "API identifier", "claude-opus-5-5" ], [ "Standard token prices", "USD $4 input and $20 output per million tokens." ], [ "Cache prices", "USD $0.20 cache reads and $5 cache writes per million tokens." ], [ "Fast-mode prices", "USD $8 input and $40 output per million tokens." ], [ "Access", "The announcement lists Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure." ], [ "Context and modality limits", "Not recorded from the launch announcement in this profile. Check the endpoint documentation before selecting a workload." ] ], "decision": "Separate token price from task cost. A coding task can require several turns, tool calls, cached context and failed attempts. Build an evaluation around a concrete acceptance condition such as passing a project test suite, then record the complete billable work needed to reach it. The launch page contains provider benchmarks; this profile does not treat those as independent DualView measurements.", "checklist": [ "Record the exact model identifier, platform and execution settings.", "Separate cache reads, cache writes and output tokens when estimating a repeated workflow.", "Compare final artifacts against the same acceptance criteria, including unsuccessful attempts." ], "limitations": "The standard rates are announcement figures, not a quote for every cloud marketplace or contract. This profile contains no measured speed, acceptance rate or benchmark ranking. Check the provider's current constraints for your use case.", "contextNotes": [ { "title": "Sonnet 5.5 launched; Haiku 5.5 remains on the roadmap", "text": "Anthropic announced Sonnet 5.5 on September 28. Its launch page still places Haiku 5.5 in the coming weeks. This replaces the earlier note that both models were upcoming. Keep each model’s pricing and platform limits separate when planning a comparison.", "source": "https://www.anthropic.com/claude-sonnet-5-5", "sourceTitle": "Anthropic Sonnet 5.5 launch" }, { "title": "How to plan a comparison across the model family", "text": "Keep an Opus baseline with saved prompts, expected outputs and acceptance criteria. When a new endpoint becomes available, verify its exact identifier and rates before replaying that workload. Record failed attempts and tool costs as well as successful answers. A lower token price would not by itself prove a lower cost per accepted task. This is an evaluation plan, not a prediction of Sonnet or Haiku performance.", "source": "https://www.anthropic.com/claude-opus-5-5", "sourceTitle": "Official announcement and pricing context" } ] }, { "slug": "glm-5-2", "name": "GLM-5.2", "provider": "Z.ai", "category": "Language model", "reviewed": "2026-09-27", "summary": "An open-weight text-generation model whose official model card documents local serving and links to hosted API access.", "source": "https://huggingface.co/zai-org/GLM-5.2", "facts": [ [ "Repository identifier", "zai-org/GLM-5.2" ], [ "Task", "Text generation" ], [ "License", "The official model card is labeled MIT." ], [ "Context", "The provider describes a 1M-token context." ], [ "Access", "Model weights on Hugging Face; model-card instructions for local serving and a link to Z.ai API services." ], [ "Pricing and release date", "Not verified in this profile; no per-token estimate or release date is assigned." ], [ "Version caveat", "The reviewed page points to a newer GLM-5.3-BF16 model. GLM-5.2 is not presented here as the latest GLM release." ] ], "decision": "Choose the deployment question before comparing scores. Running weights yourself adds hardware capacity, serving software and operations to the evaluation. A hosted endpoint has different billing and configuration constraints. Keep the exact weight revision, quantization, context settings and serving stack with your output samples; an API model name alone is insufficient to reproduce a local run.", "checklist": [ "Record the weight revision and serving configuration, including any quantization.", "Check memory and infrastructure requirements with the actual configuration rather than assuming that downloadable means laptop-sized.", "Measure total deployment cost separately from a hosted API bill and compare accepted outputs under matched criteria." ], "limitations": "The model card's context and performance descriptions are provider claims. DualView has not measured them for this profile. Local-serving instructions do not establish hardware feasibility for a particular machine. Consult the actual license files for deployment terms." }, { "slug": "gemini-omni-1-1-flash", "name": "Gemini Omni 1.1 Flash", "provider": "Google", "category": "Video model", "reviewed": "2026-09-27", "summary": "A documented video-generation and conversational-editing model. Check the input-clip limit separately from output duration and resolution.", "source": "https://ai.google.dev/gemini-api/docs/models/gemini-omni-flash", "facts": [ [ "API identifier and status", "Google lists gemini-omni-1.1-flash as stable and gemini-omni-flash-preview separately." ], [ "Documented update", "The model card says August 2026, with a page update dated August 27. This profile is not a new-release announcement." ], [ "Inputs / outputs", "Text, images and video inputs; video output. Editing and extension inputs are limited to clips up to 10 seconds." ], [ "Output formats", "The card lists 3–10-second output at 360p, 720p, 1080p or 4K, at 24 FPS. It does not establish that all modes support every combination." ], [ "Controls", "Conversational editing through the Interactions API, extension, upscaling and interpolation are documented." ], [ "Access", "Hosted Gemini API and a Google AI Studio entry point. Downloadable weights and a weight license are not established by this source." ], [ "Pricing and availability limits", "Rates, regional eligibility, quotas and any per-mode restrictions are not verified in this profile. Confirm them for the selected endpoint before use." ] ], "decision": "The useful distinction is generating a new clip versus revising an existing one. An editing workflow needs a record of the original clip, each instruction and the accepted revision; the final file alone cannot explain what changed. Before choosing this model for a longer source, check whether your intended operation accepts that source length. A supported output resolution is also not evidence of native rendering at that resolution or of preserved fine detail after an edit.", "checklist": [ "Record the exact endpoint identifier, operation, source duration, output duration, requested resolution and aspect ratio.", "For an edit, inspect unchanged regions and temporal continuity alongside the requested change. Compare the same timestamps before and after.", "Keep the original clip and the sequence of instructions with every accepted revision. Count rejected revisions when estimating cost." ], "limitations": "These are documentation-based selection notes, not measured results. DualView has not tested this model or verified its speed, audio behavior, identity consistency or visual quality. A separate technical guide may establish mode-specific details; this profile deliberately does not infer them from the general model card. No matching provenance-backed example is attached. Use saved outputs you already own for a visual review." }, { "slug": "qwen-image-2-1", "name": "Qwen-Image-2.1", "provider": "Qwen / Alibaba", "category": "Image model", "reviewed": "2026-09-28", "summary": "A downloadable image model with documented RGBA generation and editing. Evaluate transparent edges and reference fidelity separately.", "source": "https://huggingface.co/Qwen/Qwen-Image-2.1", "facts": [ [ "Identity and access", "The official Qwen/Qwen-Image-2.1 repository provides weights and Diffusers instructions. This catalog addition is not a new-release announcement." ], [ "Inputs and outputs", "The card describes text-to-image generation, image editing and transparent RGBA output." ], [ "Reference and edit controls", "Qwen documents up to 10 reference images, with circles, painted annotations or separate masks for local changes." ], [ "Documented dimensions", "Examples use 2048 × 2048 pixels; the aspect-ratio table includes landscape 2752 × 1536 and portrait 1536 × 2752. These are documented configurations, not measured quality guarantees." ], [ "Architecture qualifier", "The stated 7B parameter count applies to the visual generation component. It is not a verified memory requirement for a complete deployment." ], [ "License", "The repository identifies the Qwen Research License Agreement. Read its terms before choosing a deployment; downloadable weights do not establish unrestricted commercial permission." ], [ "Pricing and API", "Hosted API rates and availability are not verified here. Local inference still has hardware and operating costs." ] ], "decision": "For a sticker, product cutout or layered composition, a plausible subject is only part of an acceptable result. Our suggested review is to place the saved RGBA output over both a light and a dark background, then inspect fine hair, semitransparent material and boundary halos. Keep that assessment separate from whether a requested edit preserved the subject. For a multi-reference task, write down which reference supplies each attribute before reviewing the output; otherwise an attractive result can conceal a missed instruction. These are evaluation suggestions, not findings from a DualView test.", "checklist": [ "Keep the exact weight revision, inference library version, prompt, seed, step count, dimensions and input files with every output.", "Preserve the alpha channel in the saved file. A flattened preview cannot establish that the delivered asset has useful transparency.", "Review requested changes and untouched regions separately. Record rejected attempts alongside accepted outputs when estimating cost." ], "limitations": "DualView has not generated samples or measured this model’s speed, memory use, typography, identity retention or quality. The model card establishes documented capabilities, not reliability on your assets. Hardware feasibility depends on the full configuration; a parameter count alone is insufficient. No provenance-matched example or independent ranking is presented.", "contextNotes": [ { "title": "Qwen-Image 2.1 versus the 3.1 roadmap", "text": "Alibaba’s September 22, 2026 announcement describes Qwen-Image 3.1 as planned for later in 2026, with transparent backgrounds and image editing among its advertised capabilities. That is a roadmap statement, not proof of public access. The downloadable repository documented above is Qwen-Image-2.1. Do not transfer its reference-image limits, dimensions, architecture or license to 3.1. We have not verified a 3.1 checkpoint, public API, price or exact launch date.", "source": "https://www.alibabacloud.com/en/press-room/alibaba-unveils-roadmap-on-full-stack-ai-strategy?_p_lc=1", "sourceTitle": "Alibaba’s dated Qwen-Image 3.1 roadmap" }, { "title": "What this means for an evaluation today", "text": "If you need to evaluate downloadable weights now, record the exact 2.1 repository revision. If you are waiting for 3.1, treat its proposed capabilities as a future requirement to verify, not a measured improvement over 2.1. Transparency is already documented for 2.1, so the feature name alone does not establish a quality upgrade. Check alpha edges, unchanged regions and reference fidelity on the actual version you use.", "source": "https://huggingface.co/Qwen/Qwen-Image-2.1", "sourceTitle": "Qwen-Image-2.1 model card" } ] }, { "slug": "gpt-6-sol", "name": "GPT-6 Sol", "provider": "OpenAI", "category": "Language model", "reviewed": "2026-09-28", "summary": "A hosted language model for code and agent tasks. Check input limits, tool configuration and full-request billing before comparing outputs.", "source": "https://developers.openai.com/api/docs/models/gpt-6-sol.md", "facts": [ [ "API identity / access", "gpt-6-sol; hosted Responses and Chat Completions APIs. Downloadable weights are not documented." ], [ "Inputs / outputs", "Text and image inputs; text output. Image-generation tool support is separate from native output." ], [ "Token limits", "1,050,000 context; maximum input 922,000; maximum output 128,000 tokens." ], [ "Standard text-token rates", "USD per million tokens: input $2; cached input $0.20; cache writes $2.50; output $10." ], [ "Long-request adjustment", "Above 272,000 input tokens: 2× input/cache and 1.5× output rates for the full request." ], [ "Reasoning / tools", "Effort: none, low, medium (default), high, xhigh, max. Chat Completions function calling requires none; Responses supports tools." ], [ "Other billing conditions", "Regional processing adds 10% where available. Batch/Flex cost 50% of Standard; Fast costs 2× applicable rates. Tool charges are separate." ], [ "Limitations in the card", "Fine-tuning is not supported. This profile does not establish a Chat subscription entitlement or an Ultrafast rollout." ] ], "decision": "For a document or code task, decide first whether the model must read an image, return text, or call a separate tool. Those are different operations to evaluate. Preserve the original input and score the accepted answer against the same requirements for every candidate. A large context allowance does not establish that a particular answer uses all of that material reliably. When comparing a long document with a shorter extract, record both correctness and the effect of changing the input; otherwise the model comparison is also an uncontrolled comparison of prompts.", "checklist": [ "Record the API, exact model identifier, reasoning setting, tool list and saved inputs. A model name alone does not describe the execution.", "Separate uncached input, cache reads, cache writes, output and tool usage. Count failed attempts when evaluating the cost of a completed task.", "Choose a fixed acceptance condition before reading outputs. Report provider demonstrations separately from any results you actually measure." ], "limitations": "This documentation-based profile contains no measured speed, accuracy or quality ranking. The token fields are provider limits, not a guarantee that every application accepts the same request. Pricing depends on the complete request and selected processing mode; the listed rates are not a subscription quote. Weight licensing is not established because no downloadable weights are documented. Claims about future interface access belong in the separately labeled rumor report.", "relatedArticles": [ { "title": "Unverified GPT-6 Sol and Ultrafast DevDay rumor", "path": "/blog/ai-tools/gpt-6-sol-ultrafast-devday-rumor.html" } ] }, { "slug": "deepseek-v4-1-flash", "name": "DeepSeek-V4.1-Flash", "provider": "DeepSeek", "category": "Language model", "reviewed": "2026-09-28", "summary": "A multimodal API model with visual understanding. Read the current changelog before treating older V4 Pro migration notices as current routing policy.", "source": "https://api-docs.deepseek.com/updates/", "facts": [ [ "Release date", "September 10, 2026, according to the API changelog." ], [ "API name", "deepseek-flash selects V4.1-Flash in the reviewed documentation." ], [ "Modality", "The provider documents native multimodal visual understanding. This is not a claim of image or video generation." ], [ "Legacy Flash names", "The changelog says deepseek-v4-flash and deepseek-v4-flash-vision-exp temporarily route to V4.1-Flash." ], [ "V4 Pro service", "The changelog says API service continues after September 14 with unchanged billing, pending further notice." ], [ "Limits, license and rates", "Exact context/output limits, weight license and numeric prices were not verified for this profile. Check the linked official documentation before deployment." ] ], "contextNotes": [ { "title": "Why the two official notices need to be read together", "text": "The launch page still describes a plan to route V4 Pro requests to V4.1-Flash from September 14 at 04:00 UTC until V4.1-Pro arrives. That announcement is not proof that this migration happened. The API changelog instead says V4 Pro service continues in response to user demand. The revision time of that continuation paragraph is not given. Both statements were visible when checked on September 28.", "source": "https://www.deepseek.com/en/news/deepseek-v4-1-flash/", "sourceTitle": "Original launch notice" }, { "title": "Use the continuation notice for planning; verify execution separately", "text": "The operational changelog is the relevant correction to the older migration plan. It does not supply a V4.1-Pro release date. For an existing integration, verify the selected endpoint and actual response metadata before attributing an output to a model. DualView has not made API calls to test routing or account access.", "source": "https://api-docs.deepseek.com/updates/", "sourceTitle": "API changelog and continuation notice" } ], "decision": "The immediate decision is whether to retain an existing Pro integration or evaluate Flash deliberately. Do not treat a legacy alias as an immutable version pin. Save the requested model name, returned metadata, request date, inputs and tool configuration together. If two runs use different underlying versions, a change in answer quality cannot be attributed only to the prompt. For a document-reading task, include visual material that matters to the answer and inspect whether the returned explanation actually uses it; this is a proposed evaluation, not a measured result.", "checklist": [ "Separate the exact request identifier from the model identity documented at the time of the request. Retain response metadata without publishing secrets.", "Compare a fixed set of tasks and acceptance conditions. Record failures, corrections and total billed usage, not only the most attractive answer.", "Check live rates and relevant request limits before a migration. Do not copy V4 Pro prices or assume a compatibility alias keeps its previous behavior." ], "limitations": "This profile reconciles public documentation; it does not independently verify backend routing, account entitlement, latency or model quality. Provider benchmark claims are not reproduced as DualView results. The launch page links deployment resources, but this review does not establish a weight license or hardware requirement. An unchanged source recheck is not a new model release." }, { "slug": "claude-sonnet-5-5", "name": "Claude Sonnet 5.5", "provider": "Anthropic", "category": "Language model", "reviewed": "2026-09-28", "summary": "Released September 28, with a documented Google Cloud endpoint and separate provider token pricing.", "source": "https://www.anthropic.com/claude-sonnet-5-5", "facts": [ [ "Release date", "September 28, 2026" ], [ "API identifier", "claude-sonnet-5-5 (Google Cloud documentation)" ], [ "Availability", "Google Cloud lists general availability; account access and quotas still apply." ], [ "Standard provider rates", "USD $2 input and $10 output per million tokens; cache reads $0.20 per million tokens. These launch prices are not a Google Cloud contract quote." ], [ "Inputs and output", "Google Cloud: text, image and PDF inputs; text output." ], [ "Token limits", "Google Cloud lists maximum input 1,000,000 tokens and maximum output 128,000 tokens. Check request budgeting in your chosen endpoint." ], [ "Weights and license", "Hosted access is documented. No downloadable weights or open-weight license is established by the reviewed sources; service terms govern access." ] ], "contextNotes": [ { "title": "Documented endpoint, not a leaked registry", "text": "The Google Cloud reference supplies the model identifier, GA status, modalities and limits above. Earlier reports discussed different client context figures. The documentation establishes the production endpoint; it does not authenticate a private checkpoint or explain every client reservation.", "source": "https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/partner-models/claude/sonnet-5-5", "sourceTitle": "Google Cloud Sonnet 5.5 reference" } ], "decision": "Evaluate Sonnet on a repeatable task before changing a working deployment. For a bug fix, preserve the starting repository, failing test and allowed tools. For document analysis, retain the same source file and expected citations. Count retries, review time and tool costs alongside model tokens. A published token limit tells you what the endpoint accepts; it does not establish answer quality across the entire input. Anthropic positions Sonnet for bounded everyday work and Opus for more complex judgment. That is provider positioning, not a measured DualView ranking.", "checklist": [ "Verify account access, exact identifier and endpoint-specific limits before routing production traffic.", "Keep input, output and cache-read tokens separate in the budget; confirm additional billing categories with the platform.", "Compare accepted outputs under the same task constraints and retain failed attempts." ], "limitations": "DualView has not run this model or independently measured quality, latency or cost per completed task. Provider benchmarks and speed claims are not reproduced as our results. Cloud quotas, enabled tools and request limits can differ from a consumer chat product. Do not infer API limits from a model-picker label.", "relatedArticles": [ { "path": "/blog/ai-tools/claude-sonnet-5-5-checkpoint-rumor.html", "title": "Original Sonnet checkpoint rumor, updated after launch" } ] }, { "slug": "muse-video", "name": "Muse Video", "provider": "Meta Superintelligence Labs", "category": "Video model · announced preview", "reviewed": "2026-09-29", "summary": "Meta’s announced video-generation preview with native audio. This profile separates the July preview from unresolved current access and production limits.", "source": "https://ai.meta.com/blog/introducing-muse-image-muse-video-msl/", "facts": [ [ "Identity / version", "Muse Video; no numeric version or public API identifier established by the reviewed announcement." ], [ "Preview date", "July 7, 2026. This is the announcement date, not a verified general-release date." ], [ "Dated availability statement", "The July announcement says it is coming soon to creators and Meta AI. That wording does not establish September access." ], [ "Modalities", "Video generation with native audio; text prompts accompany the preview examples. A formal input schema is not supplied." ], [ "Relationship to Muse Image", "Meta says the models share a pretraining base. Image-model controls are not thereby verified for video." ], [ "Provider-described strengths", "Meta describes prompt adherence, visual fidelity and temporal consistency. These are vendor claims, not our measured results." ], [ "Provider-described limitations", "Meta identifies audio-video synchronization and physically accurate fast motion as areas needing improvement." ], [ "Pricing / license / weights", "Not established for Muse Video by the reviewed announcement. No per-second price, commercial license or downloadable weights are asserted." ], [ "Output and input limits", "Resolution, maximum duration, frame rate, reference limits and prompt length remain unverified here. Example prompts are not endpoint limits." ] ], "decision": "Treat the preview as a candidate to investigate for scenes that need both motion and sound. For a production decision, first establish a usable access route and the exact service terms. A clip that looks convincing in a provider showcase cannot tell you how reliably the model will preserve a product, follow a camera move or synchronize dialogue across repeated attempts. This is our evaluation guidance, not a performance finding.", "checklist": [ "Record the actual model identifier, access surface, date, inputs and settings before comparing outputs.", "For dialogue, inspect mouth movement and timing alongside intelligibility. For action, inspect contacts, trajectories and object continuity through the fastest motion.", "Separate prompt-only generation from reference-guided or editing tasks; include a mode only when its documentation and access are verified.", "Record accepted outputs and failed attempts before calculating cost. Do not use an empty price field as a zero-cost assumption." ], "limitations": "This profile confirms an announced identity, not general availability. We did not call the model or evaluate the preview media as a benchmark. We do not assign an Arena rank, infer a release date from third-party playground reports, or treat the Muse product family as one interchangeable model. Check current provider documentation before committing a project to a particular input mode or export format.", "relatedArticles": [ { "path": "/blog/ai-tools/muse-image-vs-seedream-5-pro-editing.html", "title": "Muse Image versus Seedream: separate still-image editing comparison" }, { "path": "/models/gemini-omni-1-1-flash.html", "title": "Gemini Omni 1.1 Flash: documented video profile" }, { "path": "/app?from=blog-video", "title": "Inspect saved video outputs in the editor" } ] }, { "slug": "minimax-m3-1-flash-preview", "name": "MiniMax-M3.1-Flash-Preview", "provider": "MiniMax", "category": "Language model · restricted preview", "reviewed": "2026-09-29", "summary": "A documented coding and reasoning preview with text, image and video input. Access is restricted to Token Plan and MiniMax Code; this is not a video generator.", "source": "https://platform.minimax.io/docs/guides/text-generation", "facts": [ [ "Identity / access", "MiniMax-M3.1-Flash-Preview. September 29 documentation restricts access to Token Plan and MiniMax Code; this review date is not a launch date." ], [ "Supported modalities", "Text, image and video input for understanding; text answers with separate thinking content." ], [ "Context window", "1,000,000 tokens, according to the provider. A context window is not a verified maximum answer length." ], [ "Reasoning controls", "Five effort levels: low, medium, high, xhigh and max; max is the default. Thinking cannot be disabled." ], [ "Request compatibility", "Anthropic-compatible and OpenAI-compatible formats are documented. Compatible format does not remove the preview access restriction." ], [ "Provider positioning", "Designed for coding, tool use, agentic reasoning and long-context tasks. These are provider descriptions, not our measured strengths." ], [ "Pricing / weights / license", "No verified unit price, downloadable weights or model-specific license established in this review. Check the current plan terms before use." ], [ "Output limit", "Not established here. The example request value is not treated as the service maximum." ] ], "decision": "The useful question is whether adjustable reasoning helps a particular coding or document task under the access conditions you actually have. Start with a task whose outcome you can verify, such as a small code change with acceptance checks or a document question with a known answer. Preserve the input, version and effort setting so a later run is comparable. This is evaluation guidance, not a recommendation based on hands-on results.", "checklist": [ "Check your account has access to this exact preview identifier before planning an integration. A compatible SDK alone does not establish entitlement.", "Separate input understanding from media generation when selecting alternatives. An image or video supplied to a language model is not evidence that it can render a new clip.", "Compare effort settings on the same task and judge the usable final answer. Record waiting time, rejected outputs and corrections if you later perform your own evaluation.", "Keep plan allowances and actual usage records alongside the task outcome. An unavailable price field must remain unknown rather than becoming a zero-cost estimate." ], "limitations": "We did not call this model, inspect paid-account access or measure its output quality. The preview designation should stay attached to saved results. No predecessor speed figure, benchmark rank, architecture or output cap is transferred to this version. Provider documentation establishes the listed interface, not reliability on every task. The Token Plan page returned no readable pricing body in this check, so this profile makes no subscription-price claim.", "relatedArticles": [ { "path": "/models/claude-sonnet-5-5.html", "title": "Claude Sonnet 5.5: separate language-model profile" }, { "path": "/models/gpt-6-sol.html", "title": "GPT-6 Sol: separate coding-model profile" }, { "path": "/blog/ai-tools/kimi-k3-1-release-rumor.html", "title": "Kimi K3.1: explicitly unverified successor report" } ] }, { "slug": "gpt-6-1-sol", "name": "GPT-6.1 Sol", "provider": "OpenAI", "category": "Language model", "reviewed": "2026-09-29", "summary": "A distinct Sol version for text and image-input tasks. Check reasoning and tool-calling compatibility before replacing an earlier endpoint.", "source": "https://developers.openai.com/api/docs/models/gpt-6.1-sol.md", "facts": [ [ "Identity and API", "gpt-6.1-sol; hosted Responses and Chat Completions. Tool calling requires Responses." ], [ "Modalities", "Text and image inputs; text output. Audio and video are unsupported; image generation is a separate tool." ], [ "Limits", "1,050,000 context tokens; maximum input 922,000; maximum output 128,000. Knowledge cutoff April 30, 2026." ], [ "Reasoning", "low, medium (default), high, xhigh, max. Neither none nor minimal is supported." ], [ "Standard token pricing", "USD per million tokens: input $2; cached input $0.10; cache writes $2.50; output $10." ], [ "Long requests", "Above 272K input tokens, the full request uses 2× input/cache and 1.5× output rates." ], [ "Processing tiers", "Fast: 2× Standard; Batch/Flex: 50% less. Regional processing: 10% premium where available." ], [ "Restrictions", "Fine-tuning unsupported. US/EU residency supported; Fast unavailable with EU residency. Downloadable weights and weight licensing are not established." ] ], "decision": "Treat a version switch as a compatibility change before treating it as a quality upgrade. A caller that previously disabled reasoning must choose a supported setting, and a Chat Completions integration that depends on tool calls needs a different route. First test request construction and tool execution using retained inputs; then compare whether the final answer meets the same acceptance criteria. This sequence helps separate an integration failure from an answer-quality difference. It is an evaluation plan, not a report of tests performed here.", "checklist": [ "Record the exact model identifier, endpoint, reasoning effort and tool configuration alongside each saved output.", "For cost comparisons, count failed attempts and tool usage as well as successful output. Separate cache reads from writes and preserve the input length.", "Use representative document and coding tasks with fixed requirements. A larger context limit does not prove that every detail in a long input will be used correctly." ], "limitations": "This profile reports provider documentation and contains no independent quality, speed or accuracy measurements. Listed token rates are not subscription charges. A supported API feature does not establish that every account, interface or region exposes it. The profile does not infer downloadable weights or an open-source license.", "contextNotes": [ { "title": "Launch access and version boundary", "text": "OpenAI announced GPT-6.1 Sol on September 29 for the API and for Work and Codex on Plus, Pro, Business, Enterprise and Edu. Its announcement explicitly excludes Chat for now and describes Sol Ultrafast as forthcoming. Keep this version separate from GPT-6 Sol and from Astra speed tiers.", "source": "https://openai.com/index/introducing-gpt-6-1-sol/", "sourceTitle": "Official launch and availability" } ], "relatedArticles": [ { "title": "DevDay rumor follow-up: confirmed details and preserved speculation", "path": "/blog/ai-tools/gpt-6-sol-ultrafast-devday-rumor.html" } ] }, { "slug": "amazon-nova-reel-v1-0", "name": "Amazon Nova Reel v1:0", "provider": "Amazon", "category": "Video model", "reviewed": "2026-09-29", "summary": "A legacy Bedrock video model with a documented September 30, 2026 end-of-life date. Check the exact version before planning a migration.", "source": "https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-amazon-nova-reel.html", "facts": [ [ "Exact identity", "amazon.nova-reel-v1:0; this profile does not transfer specifications from v1:1." ], [ "Task and modalities", "Short video generation from text and image prompts, with camera-motion controls. Video output; audio and video inputs are not supported." ], [ "Hosted API", "Amazon Bedrock runtime, using StartAsyncInvoke. The card does not support Invoke or Converse for this model." ], [ "Documented status", "Legacy; model EOL date September 30, 2026. This is a provider schedule, not a live inference check." ], [ "Listed regions", "us-east-1, eu-west-1 and ap-northeast-1, each labeled Legacy with the same EOL date." ], [ "Pricing and weights", "The card links to Bedrock pricing without quoting a rate. No current rate or downloadable-weight license is established in this review." ] ], "contextNotes": [ { "title": "What the retirement date means", "text": "AWS lists March 30, 2026 as the legacy date and September 30 as EOL for this identifier. Its policy says requests fail on or soon after EOL unless a private continued-access arrangement exists; migration is not automatic. The table also lists Nova Reel v1:1 and Nova Canvas v1:0 for September 30. They are separate identifiers, not fallback versions proven usable by this profile. No exact cutoff time or timezone is stated in the checked policy.", "source": "https://docs.aws.amazon.com/en_en/bedrock/latest/userguide/model-lifecycle-legacy.html", "sourceTitle": "AWS legacy lifecycle policy and model table" } ], "decision": "For an existing integration, locate the literal model identifier and region before choosing a replacement. A saved label such as Nova Reel may hide a version difference. Preserve the original prompt, reference image, output and acceptance criteria so a replacement can be evaluated against the job the old model served. The immediate question is continuity of that workload, not which model has the most impressive demonstration. Do not assume a name change alone preserves the request schema or delivered video.", "checklist": [ "Record the old model ID, region and invocation method from your own configuration. Keep the saved output separate from any new test result.", "Evaluate a replacement in the same lane: text-to-video or image-to-video, with matching framing and a stated motion requirement. Compare camera movement, subject continuity and task completion.", "Check replacement access, input schema, duration, resolution and billing before estimating migration cost. These are questions to verify, not promises about any alternative." ], "limitations": "No AWS generation request was made, so endpoint failure or continued availability has not been observed. This profile does not establish maximum duration, resolution, native audio output or current charges. It reports the exact provider identity and lifecycle evidence instead of borrowing specifications from another Nova version. Provider capability descriptions are not independent quality findings.", "relatedArticles": [ { "title": "H3 Max versus H3: a separate video-model comparison, not a drop-in migration guarantee", "path": "/blog/ai-tools/minimax-h3-max-vs-h3.html" } ] }, { "slug": "hy-image-3-5-preview", "name": "Hy Image 3.5 Preview", "provider": "Tencent Hunyuan", "category": "Image model · preview", "reviewed": "2026-09-29", "summary": "Tencent’s image-generation and editing preview: distinguish automatic size tiers from explicit 4K output, and regional API rates from international pricing.", "source": "https://intl.cloud.tencent.com/jp/document/product/1300/83708", "facts": [ [ "Exact identity", "hy-image-v3.5-preview on Tencent Cloud TokenHub. Preview identity retained as of September 29, 2026; no general-availability date established here." ], [ "Inputs and tasks", "Text and reference images for generation and multi-turn editing; synchronous Chat/Messages requests." ], [ "Documented limits", "100k-token prompt limit; up to 20 reference images, each at most 20 MB, in PNG/JPEG/JPG via URL or Base64." ], [ "Output sizing", "Automatic area tiers stop at 2K. Explicit size supports 4096×4096; the documented area ceiling is 16,777,216 pixels. This is an output-size claim, not proof of native 4K synthesis." ], [ "Access", "Tencent Cloud account, activated TokenHub, API key and model-specific postpaid billing are prerequisites." ], [ "Weights and license", "No downloadable weights or open-weight license established by the checked API guide. Hosted service access is not an open-weight release." ] ], "contextNotes": [ { "title": "Regional price, with units", "text": "Tencent’s mainland TokenHub product table labels these as Guangzhou-region prices: RMB 0.15 per image at 1K or 2K, and RMB 0.20 at 4K. The calculation uses RMB 10 per million tokens, with 15,000 or 20,000 tokens per image respectively. These prices were visible in the search-indexed official table; direct page extraction did not expose the table. International pricing was not independently verified and should not be inferred by currency conversion.", "source": "https://cloud.tencent.com/product/tokenhub", "sourceTitle": "Tencent TokenHub regional pricing table" }, { "title": "The size control matters", "text": "The guide distinguishes generate_max_pixels from size: use explicit size for output above the automatic 2K tier. Do not assume another application’s reference-image or output limits equal this API contract.", "source": "https://intl.cloud.tencent.com/jp/document/product/1300/83708", "sourceTitle": "Tencent Hy Image API guide" } ], "decision": "The useful evaluation is an editing sequence, not only a beautiful first image. Start with a product composition whose logo, label and geometry must survive a requested background change. Then ask for a layout adjustment without changing the product. Save both intermediate results. This tests whether the workflow can preserve what matters while changing only what you asked to change; the profile does not claim Tencent passes that test. If your job is a single poster instead, evaluate legible text, layout and export dimensions before spending time on a multi-turn integration.", "checklist": [ "Record the exact API identity, region, requested dimensions and reference count. Keep the original reference files and each edit instruction beside the corresponding output.", "Compare the same image task with an alternative you can access, such as the separately documented Qwen Image 2.1. Use identical acceptance criteria and disclose any input or size limits that prevent a matched comparison.", "Inspect small print at the intended display size, not only zoomed out. Check subject edges and unchanged areas after every edit; distinguish extra output pixels from extra detail.", "Estimate cost per accepted result from your actual invoice and rejected attempts. A low listed rate does not establish how many edits your own job will require." ], "limitations": "This is documentation research, with no generation or hands-on quality test. The reviewed sources do not establish independent text-rendering accuracy, editing reliability, inference speed, an open-weight license or an international 4K rate. The listed API limits do not certify the same limits inside every consumer application. We do not copy specifications from Hy Image 3.0, infer a release date from the catalog addition, or treat the preview label as a promise of a later launch date.", "relatedArticles": [ { "title": "Qwen Image 2.1: a separate image-model profile for evaluation planning", "path": "/models/qwen-image-2-1.html" } ] }, { "slug": "amazon-nova-canvas-v1-0", "name": "Amazon Nova Canvas v1:0", "provider": "Amazon", "category": "Image model · legacy", "reviewed": "2026-09-30", "summary": "AWS lists September 30, 2026 as this image model’s end-of-life date. Identify the exact Bedrock integration before selecting an image-generation replacement.", "source": "https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-amazon-nova-canvas.html", "facts": [ [ "Exact identity", "amazon.nova-canvas-v1:0 on Amazon Bedrock." ], [ "Inputs and output", "Text and image inputs; image output. The card does not support audio or video inputs." ], [ "Documented API", "Invoke through bedrock-runtime; not Converse, Chat Completions or Responses." ], [ "Lifecycle", "Launched December 3, 2024; card labels it Legacy with EOL September 30, 2026." ], [ "Listed regions", "us-east-1, eu-west-1 and ap-northeast-1, each carrying the same EOL label." ], [ "Pricing and license", "The card links Bedrock pricing and AWS terms. This review establishes no current per-image rate or downloadable-weight license." ] ], "contextNotes": [ { "title": "A retirement schedule, not an observed outage", "text": "AWS lists March 30, 2026 as the legacy date. Its policy says requests fail on or soon after EOL unless there is a private continued-access arrangement. Migration does not happen automatically. No exact cutoff time is stated. We made no inference request, so we cannot report whether a particular account still receives output today.", "source": "https://docs.aws.amazon.com/bedrock/latest/userguide/model-lifecycle-legacy.html", "sourceTitle": "AWS legacy lifecycle policy and model table" } ], "decision": "Start with the job Canvas served. A product-background edit and a fresh illustration are different requirements even when both produce an image. Keep a small set of your existing inputs and accepted outputs, and write down what must remain unchanged: a label, silhouette, color, layout or subject identity. Choose a replacement whose documented inputs support that same task. A convincing first sample cannot establish that the integration will handle the rest of your workload.", "checklist": [ "Read the literal model identifier and region from your configuration; a friendly model label can hide the version. Preserve the request schema and original references.", "For an editing workflow, inspect the unchanged product as closely as the changed background. Check lettering, edges and geometry at the delivery size. For text-to-image, use the same content and layout requirements.", "Confirm replacement access, output dimensions, response format and billing units before changing application routing. Compare saved outputs in the editor; do not mistake a different API contract for a drop-in replacement.", "Keep a documented fallback for failed requests and validate it on your own acceptance criteria. This profile does not certify any alternative or perform a migration." ], "limitations": "Documentation research only; no generated examples, independent quality findings or endpoint availability test. Maximum dimensions, editing-mode limits, current charges and open-weight access are not established here. Do not import specifications from Nova Reel: it is a separate video model. A model card remaining online is not proof that requests still work.", "relatedArticles": [ { "title": "Qwen Image 2.1: a separately documented image-model alternative to evaluate", "path": "/models/qwen-image-2-1.html" }, { "title": "Hy Image 3.5 Preview: image editing and access limitations", "path": "/models/hy-image-3-5-preview.html" } ] }, { "slug": "gemini-4-argon", "name": "Gemini 4 Argon", "provider": "Google DeepMind", "category": "Language model · restricted rollout", "reviewed": "2026-10-01", "summary": "A named model with restricted initial access. Separate its announced token prices and output limit from public endpoint availability.", "source": "https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/", "facts": [ [ "Identity and announcement", "Google DeepMind announced Gemini 4 Argon on September 30, 2026. A public API identifier is not established here." ], [ "Access", "Initial rollout through Fairwind to trusted cyber defenders. Broader release is forthcoming, starting with paid API customers and Google AI Ultra subscribers; no date is specified." ], [ "Provider-described uses", "Long-running software engineering, enterprise knowledge work and defensive cybersecurity; multimodal understanding is described. Exact API input formats remain unverified." ], [ "Output limit", "Google states 1 million output tokens. This is not a verified input or total context-window limit." ], [ "Announced introductory pricing", "USD $2 input and $10 output per million tokens; cached input is 95% cheaper, implying $0.10 per million cached input tokens." ], [ "After introductory pricing", "USD $4 input and $20 output per million tokens. Introductory expiry date, other charges and subsequent cached-input terms are not established." ], [ "Weights and license", "No downloadable weights or open-weight license established in this review." ] ], "decision": "The immediate decision is whether access fits your evaluation schedule. An announced model is not automatically an endpoint your account can invoke. Keep a working language model for delivery commitments and reserve an Argon comparison until you can record the exact service, version and settings. Treat provider capability descriptions as hypotheses to evaluate on your own tasks, not an independent recommendation.", "checklist": [ "For coding, prepare reproducible failing cases and regression checks. Require an inspectable change, not merely a persuasive explanation of a proposed fix.", "For document work, preserve source material and assess whether conclusions can be traced to it. Measure omissions and unsupported assertions alongside task completion.", "Budget input and output separately. As arithmetic only, 100,000 uncached input tokens plus 10,000 output tokens would cost $0.30 at the announced introductory rates or $0.60 at the stated later rates, excluding other charges. This is not a usage measurement or a current access offer.", "Keep the output ceiling separate from the amount a task needs. Set practical stopping criteria and record retries before comparing cost per completed task with an available alternative." ], "limitations": "Documentation research only; no model requests, performance measurements or independent results. Context capacity, precise modality schemas, endpoint ID, regional eligibility and general availability remain unverified. The earlier anonymous Arena reports do not establish Argon’s identity in those sessions.", "relatedArticles": [ { "title": "Gemini Arena rumors: announcement follow-up and remaining uncertainty", "path": "/blog/ai-tools/gemini-4-arena-rumor.html#argon-update" }, { "title": "GPT-6.1 Sol: a separate language-model profile for evaluation planning", "path": "/models/gpt-6-1-sol.html" }, { "title": "Claude Opus 5.5: separately documented model profile", "path": "/models/claude-opus-5-5.html" } ] } ], "changes": [ { "date": "2026-10-01", "id": "gemini-4-argon-profile", "title": "Gemini 4 Argon profile added", "description": "Recorded restricted rollout, announced token pricing and output limit; public API access remains unverified.", "path": "/models/gemini-4-argon.html" }, { "date": "2026-09-30", "id": "nova-canvas-v1-0-eol-profile", "title": "Nova Canvas v1:0 retirement profile added", "description": "Recorded exact image-model identity and AWS EOL schedule; no endpoint outage or automatic migration claimed.", "path": "/models/amazon-nova-canvas-v1-0.html" }, { "date": "2026-09-29", "id": "hy-image-3-5-preview-profile", "title": "Hy Image 3.5 Preview profile added", "description": "Documented image editing, API size controls and regional RMB pricing; international rates and native 4K synthesis remain unverified.", "path": "/models/hy-image-3-5-preview.html" }, { "date": "2026-09-29", "id": "nova-reel-v1-0-eol-profile", "title": "Nova Reel v1:0 retirement profile added", "description": "Recorded AWS September 30 EOL schedule and exact model/API identity. No live endpoint test or automatic replacement is claimed.", "path": "/models/amazon-nova-reel-v1-0.html" }, { "date": "2026-09-29", "id": "gpt-6-1-sol-profile", "title": "GPT-6.1 Sol profile added", "description": "New version profile covers reasoning compatibility, tool-call routing, documented limits and token charges. Launch access is distinguished from forthcoming Ultrafast.", "path": "/models/gpt-6-1-sol.html" }, { "date": "2026-09-29", "id": "minimax-m3-1-flash-preview-profile", "title": "MiniMax M3.1 Flash Preview profile added", "description": "Documented restricted access, input modalities and reasoning controls. This is a catalog addition, not a new release announcement.", "path": "/models/minimax-m3-1-flash-preview.html" }, { "id": "muse-video-preview-profile", "date": "2026-09-29", "path": "/models/muse-video.html", "title": "Muse Video preview profile added", "description": "Added the verified July preview identity, attributed native-audio claims and limitations, and explicit access and pricing unknowns. This catalog addition is not a new release." }, { "id": "2026-09-28-sonnet-5-5-launch", "date": "2026-09-28", "title": "Claude Sonnet 5.5 launch and verified profile", "description": "Added launch pricing and Google Cloud endpoint facts; corrected the Opus family roadmap and linked the updated rumor record.", "path": "/models/claude-sonnet-5-5.html" }, { "date": "2026-09-28", "id": "deepseek-v4-1-flash-routing-profile", "title": "DeepSeek-V4.1-Flash profile: reconcile V4 Pro routing notices", "description": "Adds API identity, legacy Flash aliases and the V4 Pro continuation notice, distinguishing the changelog from the older migration plan. Catalog addition, not a new release.", "path": "/models/deepseek-v4-1-flash.html" }, { "id": "claude-5-5-family-roadmap", "date": "2026-09-28", "path": "/models/claude-opus-5-5.html", "title": "Claude 5.5 family: separate Opus access from the Sonnet and Haiku roadmap", "description": "Added the dated official roadmap, missing release and pricing details, and a practical evaluation plan without transferring Opus specifications." }, { "id": "qwen-image-2-1-roadmap-clarification", "date": "2026-09-28", "path": "/models/qwen-image-2-1.html", "title": "Qwen-Image: available 2.1 weights versus the 3.1 roadmap", "description": "Clarified Alibaba’s dated 3.1 announcement, access unknowns and why 2.1 specifications must not be copied to the planned version." }, { "date": "2026-09-28", "id": "gpt-6-sol-profile", "title": "GPT-6 Sol profile added", "description": "Documented API identity, token limits, rate adjustments and evaluation notes. This catalog addition is not a new model release.", "path": "/models/gpt-6-sol.html" }, { "date": "2026-09-27", "id": "qwen-image-2-1-profile", "title": "Qwen-Image-2.1 image profile added", "description": "Documented transparency, reference controls and evaluation notes, with explicit licensing and pricing limits. This is a catalog addition, not a release announcement.", "path": "/models/qwen-image-2-1.html" }, { "date": "2026-09-27", "id": "gemini-omni-1-1-flash-profile", "title": "Gemini Omni 1.1 Flash video profile added", "description": "Documented video input limits, output formats and editing controls, with explicit pricing and availability unknowns. This catalog addition does not announce a new release.", "path": "/models/gemini-omni-1-1-flash.html" }, { "date": "2026-09-27", "id": "initial-profiles", "title": "Three source-linked model profiles added", "description": "Initial profiles for GPT-6 Astra, Claude Opus 5.5 and GLM-5.2. This is a catalog publication event, not three new model releases.", "path": "/models.html" } ] }