Reference edition checked November 5, 2026 (ET). Reporting feeds update independently; reference facts retain their stated dates.
This reference covers twelve selected records across GPT, Claude, Gemini, Grok, DeepSeek, Muse, Llama, Mistral, Phi and Nova. It mixes current catalogue entries with explicitly labeled historical releases so readers can distinguish product families, API aliases and downloadable weights. Inclusion is not a recommendation and row order is not a ranking. All capabilities and commercial specifications are attributed to the linked developer documentation.
The original three homepage names—Gemini 4 Carbon, Muse Wonder and DeepSeek-V4.1-Flash-1017—were checked again against developer evidence. Carbon is retained as an announcement-only record. Argon specifications are not transferred to Carbon, and the 1017 suffix is not interpreted as a release date.
This is a static reference table and record register, not a searchable database. “Not recorded” means the field was not established from the documents used here, not that a model lacks that feature. No benchmark scores are pooled into an overall intelligence score. See benchmark and ranking methodology.
Swipe or scroll sideways to see all columns.
| Model / developer | Context tokens | Published dollar input / output per million tokens | Release evidence |
|---|---|---|---|
| GPT-6 Astra OpenAI | 1.05M (vendor label) | $10 / $50. Listed base input/output rates; caches, tools, batch and special tiers excluded. | Exact release date not recorded |
| Claude Opus 5.5 Anthropic | 1M (vendor label) | $4 / $20. Listed base rates, not batch/cache or partner-platform pricing. | Exact release date not recorded |
| Gemini 4 Argon Google DeepMind | 1,048,576 | Not recorded | Exact release date not recorded |
| Gemini 4 Carbon Google DeepMind | Not recorded | Not recorded | 2026-10 (announcement month) |
| Grok 4.7 xAI / SpaceXSI | 500K (vendor label) | $2 / $6. Prompts below 200k tokens. At or above 200k prompt tokens all input/output tokens use $4/$12 per million. Cached input $0.50/$1 respectively. | 2026-09-21 (company timeline release) |
| DeepSeek-V4.1-Flash-1017 DeepSeek | 1M (vendor label) | Off-peak cache-miss input $0.15 / output $0.60; peak $0.30 / $1.20. Schedule, cache and tool charges differ. | Exact release date not recorded |
| Muse Wonder Meta | Not recorded | Not recorded | 2026-10-12 (announcement listing) |
| Llama 3.3 70B Instruct Meta | 128K (vendor label) | Not recorded | 2024-12-06 (model release) |
| Mistral Large 4 Mistral AI | 1M (vendor label) | $0.7 / $2.1. Listed input/output API rates; other billing modes not included. | 2026-10-06 (model page date) |
| Mistral Small 4 Mistral AI | 256K (vendor label) | Not recorded | 2026-03-16 (model page date) |
| Phi-4-mini-instruct Microsoft | 128K (vendor label) | Not recorded | 2025-02 (model-card release month) |
| Nova 2 Lite Amazon | Not recorded | Not recorded | Exact release date not recorded |
How to compare these specifications
Tokenizers differ between developers, so the same document may consume different token counts. A model’s maximum context is not a guarantee that it will use every part reliably. Google explicitly describes Argon’s limit as a unified input-plus-output budget. Output ceilings, synchronous limits, long-decode options and reasoning budgets may impose additional constraints. A context comparison therefore needs the exact API and operating mode as well as the headline number.
Dollar rates above are the vendor-listed figures captured on the check date. They are not quotes for a complete task and do not include every cache, batch, platform, tool or tax adjustment. Grok’s long-prompt tier applies to all tokens in a qualifying request; DeepSeek’s off-peak and peak rates depend on its billing schedule. API token prices cannot be compared directly with a chatbot subscription or self-hosting hardware expense.
Model records and primary documentation
GPT-6 Astra
Text/image input; text output; reasoning and coding. Catalogue snapshot; release day and downloadable-weight license not established here.
Access: OpenAI API / Responses API Weights/license: License terms and weight availability not established in this record; API listing alone does not establish either.
Claude Opus 5.5
Text/image input; text output; adaptive reasoning, coding and tool use. Pinned model ID claude-opus-5-5. No release day inferred from retirement commitments.
Access: Claude API; catalogue also lists partner-cloud IDs Weights/license: License terms and weight availability not established in this record; API listing alone does not establish either.
Gemini 4 Argon
Text/image/video/audio/PDF inputs; text output; reasoning and tool use. Unified input-plus-output context. Page says latest update October 2026; this is not an exact launch date.
Access: Gemini API; stable model code gemini-4-argon Weights/license: License terms and weight availability not established in this record; API listing alone does not establish either.
Gemini 4 Carbon
Developer announcement and performance overview. Announcement-only record. Do not copy Argon context, pricing or access onto Carbon.
Access: Not verified in this record. Weights/license: License terms and weight availability not established in this record; API listing alone does not establish either.
Grok 4.7
Reasoning, chat and code API model. Release date is separately supported by company timeline: September 21, 2026.
Access: Developer API Weights/license: License terms and weight availability not established in this record; API listing alone does not establish either.
DeepSeek-V4.1-Flash-1017
Thinking/non-thinking modes; vision, JSON output, tool calls and APIs documented. 1017 is part of the version string, not a release date. Do not conflate API availability with weight/license availability.
Access: deepseek-flash API alias Weights/license: License terms and weight availability not established in this record; API listing alone does not establish either.
Muse Wonder
Personal and professional agentic tasks, as described by Meta. Official AI page confirms name and announcement date. Context/pricing/license not verified here.
Access: Not verified in this record. Weights/license: License terms and weight availability not established in this record; API listing alone does not establish either.
Llama 3.3 70B Instruct
Historical text-in/text-out multilingual instruction model. Historical reference, not claimed to be Meta’s latest release. Hosting prices vary by provider and are not zero.
Access: Downloadable weights via Meta distribution Weights/license: Llama 3.3 Community License (custom commercial terms and acceptable-use policy; not labeled open source)
Mistral Large 4
General-purpose open-weight multimodal model.
Access: Mistral developer API and open-weight distribution Weights/license: Mistral Ecosystem Open Weight License; not labeled open source
Mistral Small 4
Hybrid instruction, reasoning and coding model. Price labels were ambiguous in extracted model page, so rates deliberately not transcribed.
Access: Mistral developer API and open-weight distribution Weights/license: Apache 2.0, as labeled in developer catalogue
Phi-4-mini-instruct
Historical lightweight text instruction/reasoning model. Historical reference. Training dates and data cutoff are not release dates. Not claimed to be the newest Phi model.
Access: Downloadable weights; model card also references Azure Weights/license: MIT license
Nova 2 Lite
Reasoning model in Amazon Nova portfolio; a customized variant powers Nova Act browser workflows. Portfolio-level record. Exact context/pricing/release/license not transcribed without model-specific billing documentation.
Access: Not verified in this record. Weights/license: License terms and weight availability not established in this record; API listing alone does not establish either.
Open weights are not automatically open source
The Llama 3.3 record uses a custom Community License and acceptable-use policy. Mistral Large 4 is labeled with a separate Ecosystem Open Weight License, whereas Small 4 is labeled Apache 2.0. Microsoft’s Phi card states MIT. These distinctions matter for redistribution and commercial deployment. This register names the observed license rather than issuing a legal conclusion; read the actual terms for the exact model version and planned use.
Related TBR reporting
- OpenAI Launches GPT-6.1 Terra, a Cheaper AI Model Built to Operate Computers 2026-10-28
- Mistral Large 4 Preview Brings 1 Trillion Open-Weight Model 2026-10-07
- Google Unveils Gemini 4 Carbon, a Cybersecurity-Focused AI Model Most People Can’t Use Yet 2026-10-29
From the central AI news archive
Google’s Health AI Framework Puts Expert Review Beside Benchmarks
MIT Robot Memory Check Lifts Test Accuracy by 5 Percentage Points
207,049 Claude Research Sessions Show Analysis Ahead of Ideas
OpenAI Says GPT-6 Cyber Lifts ExploitGym Score to 48.5%
Moonshot AI Sought Nvidia Blackwell Access via Europe, Sources Say