SQ Magazine The Tech Index AI Model Tracker

AI Model Tracker

A curated timeline of the models that matter: who shipped what, the specs and license you can actually build on, and what it costs.

41 models tracked verified through 26 Sep 2026 last verified 12 Oct 2026 next review by 26 Oct 2026

Models tracked: 41
Open weights: 8
Developers: 11
Re-verify cycle: 30day

At a glance

Largest context window: GPT-6.1 Sol 1.1M tokens, tied with 4 other models
Most recent release: Claude Haiku 5.5 Oct 2026

All records

41 models, newest release first. Filter or search below.

How to use this tracker

Search & filter
Filter by developer, license or modality, or use the search box to find a specific model.
Explore the data
Sort by newest first, A to Z or largest context.
Inspect records
Select Detail on any row to open its full record, then copy a link that opens that row directly.
Verify sources
Every row links the primary source it came from, most often huggingface.co.

Model	Developer	Released	Modality	License	Price	Access	Our coverage
Claude Haiku 5.5	Anthropic	Oct 2026	Text + Vision	Proprietary	Low tier	API	Anthropic release
Context window 1M tokens Parameters Undisclosed Knowledge cutoff Jun 2026 Lifecycle Current Stated use High-volume, latency-sensitive tasks such as classification, extraction and routing Benchmarks (developer-reported) OSWorld 2.1 (offline) 72.4% · Terminal-Bench 4.0 39.2% · GDPval-AA v2.1 1620 Elo License terms Commercial API Primary source Anthropic release ↗ Last confirmed 8 Oct 2026 Recent change Released 7 Oct 2026 from $0.10/$0.50 per MTok ($0.50/$2.50 above 100K-token prompts); 1M context, 128K output, Jun 2026 cutoff. Haiku 4.5 moved to legacy. Security & safety System card published; cybersecurity safeguards are stricter than on Haiku 4.5.
Claude Sonnet 5.5	Anthropic	Sep 2026	Text + Vision	Proprietary	Mid tier	API	Anthropic release
Context window 1M tokens Parameters Undisclosed Knowledge cutoff Jun 2026 Lifecycle Current Stated use Best combination of speed and intelligence Benchmarks (developer-reported) Terminal-Bench 4.0 70.6% · CursorBench 4.0 55.5% · GDPval-AA v2.1 1844 Elo License terms Commercial API Primary source Anthropic release ↗ Last confirmed 2 Oct 2026 Recent change Released 28 Sep 2026 at $2/$10 per MTok, the same as Sonnet 5; 1M context, 128K output, Jun 2026 cutoff. Sonnet 5 moved to legacy. Security & safety System card published; first Sonnet model to launch with cyber safeguards and fallbacks.
GPT-6.1 Sol	OpenAI	Sep 2026	Text + Vision	Proprietary	Mid tier	API	OpenAI release
Context window 1.1M tokens Parameters Undisclosed Knowledge cutoff Apr 2026 Lifecycle Current Stated use Complex coding and professional work License terms Commercial API; OpenAI terms Primary source OpenAI release ↗ Last confirmed 30 Sep 2026 Recent change Released 29 Sep 2026 per the OpenAI API changelog. Model page: 1,050,000 context, 128,000 max output, 30 Apr 2026 cutoff, $2 in and $10 out per MTok, text and image in, text out.
Claude Opus 5.5	Anthropic	Sep 2026	Text + Vision	Proprietary	High tier	API	Anthropic release
Context window 1M tokens Parameters Undisclosed Knowledge cutoff Jun 2026 Lifecycle Current Stated use Long-running agentic coding and knowledge work Benchmarks (developer-reported) Terminal-Bench 4.0 66.4% · CursorBench 4.0 57.8% · GDPval-AA v2.1 1846 Elo License terms Commercial API Primary source Anthropic release ↗ Last confirmed 26 Sep 2026 Recent change Released 22 Sep 2026 at $4/$20 per MTok, 20% below Opus 5; 1M context, 128K output, Jun 2026 cutoff. Opus 5 and Opus 4.8 moved to legacy. Security & safety System card published.
GPT-6 Luna	OpenAI	Sep 2026	Text + Vision	Proprietary	Low tier	API	OpenAI release
Context window 1.1M tokens Parameters Undisclosed Knowledge cutoff May 2026 Lifecycle Current Stated use Focused, high-volume tasks License terms Commercial API; OpenAI terms Primary source OpenAI release ↗ Last confirmed 26 Sep 2026 Recent change Released 22 Sep 2026 per the OpenAI API changelog. Model page: 1,050,000 context, 128,000 max output, 18 May 2026 cutoff, $0.10 in and $0.50 out per MTok, text and image in, text out.
GPT-6 Sol	OpenAI	Sep 2026	Text + Vision	Proprietary	Mid tier	API	OpenAI release
Context window 1.1M tokens Parameters Undisclosed Knowledge cutoff Apr 2026 Lifecycle Current Stated use Complex coding and agentic workflows License terms Commercial API; OpenAI terms Primary source OpenAI release ↗ Last confirmed 26 Sep 2026 Recent change Released 22 Sep 2026 per the OpenAI API changelog. Model page: 1,050,000 context, 128,000 max output, 20 Apr 2026 cutoff, $2 in and $10 out per MTok, text and image in, text out.
Grok 4.7	xAI	Sep 2026	Text + Vision	Proprietary	Mid tier	API	xAI announcement
Context window 500K tokens Parameters Undisclosed Knowledge cutoff May 2026 Lifecycle Current Stated use Coding and knowledge work Benchmarks (developer-reported) CursorBench 4.0 46.3% · DeepSWE v1.1 71.0% · EEBench 64.0% License terms Commercial API Primary source xAI announcement ↗ Last confirmed 26 Sep 2026 Recent change Announced 21 Sep 2026 at 2 US dollars in and 6 out per million tokens. The xAI model docs list grok-4.7 at 500,000 tokens, text and image in, text out. No cutoff or parameter count is stated. Security & safety New safeguard stack; xAI reports HackerBench v0.3 lets 3.3% of risky dual-use prompts through and LatchBio biosafety 62.4%.
GPT-6 Astra	OpenAI	Sep 2026	Text + Vision	Proprietary	High tier	API	OpenAI Releases GPT-6 Astra After Largest Training Run Yet
Context window 1.1M tokens Parameters Undisclosed Knowledge cutoff Apr 2026 Lifecycle Current Stated use Hardest end-to-end work across computer use, browsing and software engineering License terms Commercial API; OpenAI terms Primary source OpenAI release ↗ Last confirmed 7 Oct 2026 Recent change Access set to API: OpenAI's model docs (7 Oct 2026) list standard usage-tier rate limits with no Trusted Access requirement. Context, output, cutoff and price unchanged.
Gemini 3.8 Flash	Google	Sep 2026	Multimodal	Proprietary	Low tier	API	Google announcement
Context window 1M tokens Parameters Undisclosed Knowledge cutoff Mar 2026 Lifecycle Current Stated use Long-horizon software engineering, autonomous agents, and complex enterprise workflows Benchmarks (developer-reported) HLE-Verified 54.9% · DeepSWE v1.1 73.7% License terms Commercial API; Google terms Primary source Google announcement ↗ Last confirmed 26 Sep 2026 Recent change GA 2 Sep 2026. 1,048,576 in, 65,536 out. The model card states a Mar 2026 cutoff and DeepSWE v1.1 73.7%, reversing the 3 Sep removal; HLE-Verified 54.9% is on both pages. Security & safety Model card published for Gemini 3.8 Flash.
Claude Fable 5.1	Anthropic	Sep 2026	Text + Vision	Proprietary	High tier	API	Anthropic release
Context window 1M tokens Parameters Undisclosed Knowledge cutoff Jun 2026 Lifecycle Current Stated use Demanding reasoning and long-horizon agentic work Benchmarks (developer-reported) CursorBench 3.2.0 73.4% · Terminal-Bench 4.0 55.8% · OSWorld 2.0 77.9% (partial) License terms Commercial API; Claude Mythos 5.1 shares its specs and pricing, by invitation only Primary source Anthropic release ↗ Last confirmed 26 Sep 2026 Recent change Announcement states Fable 5.1 is generally available and priced as Fable 5 was, $10/1M input and $50/1M output, with cache reads cut 75% to $0.25/1M Security & safety System card published for Fable 5.1 and Mythos 5.1; Anthropic states the benchmark scores were produced with production safeguards enabled.
Muse-Glimmer-30B	Meta	Aug 2026	Text + Vision	Open weights	Self-host	Open weights	Meta model card
Context window 128K tokens Parameters ~29.6B Knowledge cutoff Jan 2026 Lifecycle Current Stated use Multimodal agentic model sized for local deployment Benchmarks (developer-reported) SWE-Bench Pro 51.2 · AIME 2026 94.7 · MCP Atlas 75.5 · Charxiv Reasoning 78.8 License terms Apache 2.0 covering the repository and the weights, per the model card. Primary source Meta model card ↗ Last confirmed 27 Sep 2026 Recent change Re-verified 27 Sep 2026 against the model card: security note corrected, since the card does carry Trust and Safety and Preparedness sections with evaluations. All other fields held. Security & safety Open weights: you own the data boundary and the safety tuning. The card reports content-safety, agentic-risk, privacy and preparedness evaluations, rates chem/bio risk moderate or lower, and recommends deploying it behind added guardrails.
DeepSeek-V4-Flash-Vision-Exp	DeepSeek	Aug 2026	Multimodal	Open weights	Self-host	Open weights	DeepSeek model card
Context window 1M tokens Parameters 305B Lifecycle Retired Stated use Experimental open-weights multimodal agent work across image and text inputs Benchmarks (developer-reported) Terminal Bench 2.1 83.9 · DeepSWE 59.3 · Cybergym 75.3 · Chartography 64.3 License terms MIT License covering the repository, per the model card. The card states no additional commercial restriction. Primary source DeepSeek model card ↗ Last confirmed 8 Oct 2026 Recent change DeepSeek release notes of 10 Sep 2026 state V4-Flash-Vision-Exp is retired and the API now routes deepseek-v4-flash-vision-exp to V4.1-Flash. Weights stay public under MIT. Security & safety Weights public and ungated on Hugging Face across 48 safetensors shards, so self-hosting keeps data in your own boundary. The repository ships no system card or safety evaluation, and the card marks the model experimental.
Qwen3.8-Flash-Next	Alibaba	Aug 2026	Multimodal	Community	Self-host	Open weights	Qwen model card
Context window 256K tokens Parameters 125B (6B active) Lifecycle Current Stated use Long-context agentic coding and computer use at 6B active parameters Benchmarks (developer-reported) GPQA Diamond 91.7 · SWE-bench Pro 62.5 · LiveCodeBench v6 91.9 · HLE 35.9 License terms Qwen Community License 1.0. Commercial use allowed; name display above 100M MAU or USD 20M monthly revenue; separate license for MaaS or AI assistants. Primary source Qwen model card ↗ Last confirmed 27 Sep 2026 Recent change Second verification 4 Sep 2026 from the cited repo: text_config.max_position_embeddings 262,144 confirms the 256K context; model_type is qwen4_exp and HF licenses it as other, matching the experiment… Security & safety Weights public and ungated on Hugging Face across 131 safetensors shards; self-host data boundary. Unlike the Qwen3.8-Max License on the 2.4T row, this one needs a separate Qwen license for Model-as-a-Service or AI Work Assistant use with no revenue floor.
Qwen3.8-27B	Alibaba	Aug 2026	Multimodal	Open weights	Self-host	Open weights	Qwen model card
Context window 256K tokens Parameters 27B (dense) Lifecycle Current Stated use Agentic coding and computer use in a compact dense model Benchmarks (developer-reported) GPQA Diamond 89.2 · SWE-bench Pro 61.7 · Terminal Bench 2.1 73.0 · OSWorld-Verified 84.3 License terms Apache-2.0 Primary source Qwen model card ↗ Last confirmed 27 Sep 2026 Recent change Second verification 4 Sep 2026 from the cited repo: text_config.max_position_embeddings 262,144 confirms the 256K context, HF metadata gives license apache-2.0. Published. Security & safety Apache-2.0 weights, public and ungated on Hugging Face across 18 safetensors shards; self-host data boundary. The repo LICENSE is unmodified Apache-2.0 and adds no scale caps, MAU thresholds or revenue thresholds, unlike the Qwen3.8-Max License on the 2.4T row.
Gemini 3.7 Flash	Google	Aug 2026	Multimodal	Proprietary	Low tier	API	Google announcement
Context window 1M tokens Parameters Undisclosed Knowledge cutoff Mar 2026 Lifecycle Superseded Stated use Complex coding, agentic workflows, and multi-step execution Benchmarks (developer-reported) FrontierCode 1.1 Main 43.6% · DeepSWE v1.1 65.3% · AutomationBench 30.4% License terms Commercial API; Google terms Primary source Google announcement ↗ Last confirmed 26 Sep 2026 Recent change Announced 13 Aug 2026. 1,048,576 in, 65,536 out; the model card gives a Mar 2026 cutoff. $0.75/$3.75 per M to 31 Dec 2026, then $1.50/$7.50. Superseded by Gemini 3.8 Flash 2 Sep 2026.
Grok 4.6	xAI	Aug 2026	Text + Vision	Proprietary	Mid tier	API	xAI developer docs
Context window 500K tokens Parameters Undisclosed Knowledge cutoff Feb 2026 Lifecycle Current Stated use Coding, agentic tasks, and knowledge work Benchmarks (developer-reported) CursorBench v3.2 69.9% · DeepSWE v1.1 65.9% · FrontierCode v1.1 61.3% License terms Commercial API Primary source xAI developer docs ↗ Last confirmed 7 Oct 2026 Recent change xAI developer docs state a 500,000-token context window and $2 input, $6 output per million tokens.
Qwen3.8-2.4T-A95B	Alibaba	Aug 2026	Text	Community	Self-host	Open weights	Qwen model card
Context window 256K tokens Parameters 2.4T (95B active) Lifecycle Current Stated use Long-context reasoning and agentic coding Benchmarks (developer-reported) GPQA Diamond 92.6 · SWE-bench Pro 67.7 · Terminal Bench 2.1 86.6 · PaperBench 93.0 License terms Qwen3.8-Max License. Commercial use allowed; name display required above 100M MAU or $20M monthly revenue; separate license for MaaS above $50M/12 months. Primary source Qwen model card ↗ Last confirmed 26 Sep 2026 Recent change Cited repo config.json gives max_position_embeddings 262,144, confirming the 256K context Security & safety Full weights public and ungated on Hugging Face across 213 safetensors shards; self-host data boundary. The Qwen3.8-Max License requires a separate license from Qwen for Model-as-a-Service use above $50M revenue in 12 consecutive months.
DeepSeek-V4-Pro	DeepSeek	Aug 2026	Text	Open weights	Low tier	API + weights	DeepSeek model card
Context window 1M tokens Parameters 1.65T (MoE) Lifecycle Deprecated Stated use Frontier-scale open-weights MoE with 1M context at low hosted-API prices Benchmarks (developer-reported) Terminal Bench 2.1 87.9% · DeepSWE 62.7% · Cybergym 83.3% License terms MIT License covering both the repository and the model weights, per the model card. Primary source DeepSeek model card ↗ Last confirmed 8 Oct 2026 Recent change DeepSeek release notes of 10 Sep 2026 announced a V4-Pro phase-out, with API requests routed to V4.1-Flash from 14 Sep. The DeepSeek pricing page now states that API service for V4 Pro continues afte… Security & safety MIT weights allow self-hosting for full data control; review the DeepSeek provider data policy before using the hosted API. The repository ships no system card or safety evaluation beside the weights.
DeepSeek-V4-Flash	DeepSeek	Jul 2026	Text	Open weights	Self-host	Open weights	DeepSeek model card
Context window 1M tokens Parameters 304B (MoE) Lifecycle Retired Stated use Low-cost 1M-context open-weights MoE for agentic and long-context work Benchmarks (developer-reported) Terminal Bench 2.1 82.7% · DeepSWE 54.4% · Cybergym 76.7% License terms MIT License covering both the repository and the model weights, per the model card. Primary source DeepSeek model card ↗ Last confirmed 8 Oct 2026 Recent change DeepSeek release notes of 10 Sep 2026 state V4-Flash is retired and the API now routes deepseek-v4-flash to V4.1-Flash. Weights stay public under MIT on Hugging Face. Security & safety MIT weights allow self-hosting for full data control; review the DeepSeek provider data policy before using the hosted API. The repository ships no system card or safety evaluation beside the weights.
Grok 4.5	xAI	Jul 2026	Text + Vision	Proprietary	Mid tier	API	xAI developer docs
Context window 500K tokens Parameters Undisclosed Lifecycle Superseded Stated use Coding, agentic tasks, and knowledge work Benchmarks (developer-reported) SWE Marathon pass@1 29.0% · Terminal Bench 2.1 83.3% · SWE-Bench Pro 64.7% License terms Commercial API Primary source xAI developer docs ↗ Last confirmed 7 Oct 2026 Recent change xAI developer docs state a 500,000-token context window and $2 input, $6 output per million tokens.
Claude Opus 5	Anthropic	Jul 2026	Text + Vision	Proprietary	High tier	API	Anthropic release
Context window 1M tokens Parameters Undisclosed Knowledge cutoff May 2026 Lifecycle Superseded Stated use Complex agentic coding and enterprise work Benchmarks (developer-reported) OSWorld 2.0 70.6% · DeepSWE v1.1 68.8% License terms Commercial API Primary source Anthropic release ↗ Last confirmed 8 Oct 2026 Recent change SWE-bench Verified 96.0% removed: the cited launch post does not state it (its chart shows SWE-bench Pro). Security & safety System card published. Cyber classifiers allow source-code vulnerability discovery but block binary scanning, penetration testing, and exploit generation.
Kimi K3	Moonshot	Jul 2026	Multimodal	Community	Mid tier	API + weights	Moonshot’s Kimi K3 Beats Top U.S. AI Models in Blind Tests
Context window 1M tokens Parameters 2.8T (104B active) Lifecycle Current Stated use Long-context reasoning, agentic coding Benchmarks (developer-reported) Terminal-Bench 2.1 88.3% · DeepSWE 67.5% · SWE Marathon 42.0% License terms Kimi K3 License: MIT-style; MaaS over 20M USD a year needs a deal; credit Kimi K3 above 100M MAU or 20M USD a month Primary source Moonshot model card ↗ Last confirmed 8 Oct 2026 Recent change HF commit history for the weights starts 27 Jul 2026 (repo created 13 Jun). Specs and license re-checked on the model card, 8 Oct 2026. Security & safety Full weights published on Hugging Face under the Kimi K3 License. First-party Kimi API also live.
Gemini 3.5 Flash-Lite	Google	Jul 2026	Multimodal	Proprietary	Low tier	API	Google model card
Context window 1M tokens Parameters Undisclosed Knowledge cutoff Mar 2026 Lifecycle Current Stated use Cheapest, low-latency, high throughput Benchmarks (developer-reported) SWE-Bench Pro 54.2% · OSWorld-Verified 74.0% · Terminal-Bench 2.1 54.0% License terms Commercial API Primary source Google model card ↗ Last confirmed 26 Sep 2026 Recent change Current confirmed: Google lists Gemini 3.5 Flash-Lite as Stable on the models page, and its deprecations row announces no shutdown date Security & safety Model card; safety evals.
Gemini 3.6 Flash	Google	Jul 2026	Multimodal	Proprietary	Low tier	API	Google model card
Context window 1M tokens Parameters Undisclosed Knowledge cutoff Mar 2026 Lifecycle Superseded Stated use Fast workhorse, up to 17% fewer tokens Benchmarks (developer-reported) SWE-Bench Pro 58.7% · OSWorld-Verified 83.0% · Terminal-Bench 2.1 78.0% License terms Commercial API; Google terms Primary source Google model card ↗ Last confirmed 26 Sep 2026 Recent change Launch post priced it at $1.50/$7.50 per 1M; the pricing page now lists $0.75/$3.75 through 31 Dec 2026, then $1.50/$7.50. Security & safety Model card and safety evals.
GPT-5.6 Sol	OpenAI	Jul 2026	Text + Vision	Proprietary	High tier	API	OpenAI Launches GPT 5.6 Sol With Powerful New AI Features
Context window 1.1M tokens Parameters Undisclosed Knowledge cutoff Feb 2026 Lifecycle Current Stated use Flagship of the GPT-5.6 series (Sol/Terra/Luna) Benchmarks (developer-reported) Terminal-Bench 2.1 88.8% License terms Commercial API; OpenAI terms Primary source OpenAI release ↗ Last confirmed 12 Oct 2026 Recent change Model page states a 1,050,000 context window, 128,000 max output tokens and a 16 Feb 2026 knowledge cutoff Security & safety System card published; assessed under the Preparedness Framework.
Claude Sonnet 5	Anthropic	Jun 2026	Text + Vision	Proprietary	Mid tier	API	Anthropic release
Context window 1M tokens Parameters Undisclosed Knowledge cutoff Jan 2026 Lifecycle Superseded Stated use Most agentic Sonnet; near-Opus at lower cost Benchmarks (developer-reported) OSWorld-Verified 81.2% · Terminal-Bench 2.1 80.4% License terms Commercial API; Anthropic usage policy Primary source Anthropic release ↗ Last confirmed 9 Oct 2026 Recent change Enterprise data controls removed from security: not on the launch post. SWE-bench Verified 85.2% removed earlier: not on the post (chart shows SWE-bench Pro). Security & safety System card published; cyber safeguards enabled by default.
Claude Fable 5	Anthropic	Jun 2026	Text + Vision	Proprietary	High tier	API	Claude Fable 5 Ends Free Access For Pro Subscribers
Context window 1M tokens Parameters Undisclosed Knowledge cutoff Jan 2026 Lifecycle Superseded Stated use Frontier intelligence for long-running agents License terms Commercial API; Mythos-class 30-day data retention, not used for training Primary source Anthropic release ↗ Last confirmed 26 Sep 2026 Recent change Legacy since Claude Fable 5.1 (1 Sep 2026); still available at $10/$50 per MTok, retirement not before 9 Jun 2027. Security & safety System card published.
Claude Opus 4.8	Anthropic	May 2026	Text + Vision	Proprietary	High tier	API	Anthropic Unveils Claude Science to Transform Research
Context window 1M tokens Parameters Undisclosed Knowledge cutoff Jan 2026 Lifecycle Superseded Stated use Frontier reasoning, long-horizon agents Benchmarks (developer-reported) OSWorld-Verified 83.4% License terms Commercial API; Anthropic usage policy Primary source Anthropic model card ↗ Last confirmed 9 Oct 2026 Recent change GPQA Diamond 93.6% and SWE-bench Verified 88.6% removed: neither is on the cited launch post or its chart. API training-data line removed: not on the post. Security & safety System card and Responsible Scaling Policy published.
Gemini 3 Pro	Google	Nov 2025	Multimodal	Proprietary	not available	API	Google, Gemini API deprecations
Context window 1M tokens Parameters Undisclosed Knowledge cutoff Jan 2025 Lifecycle Retired Stated use Frontier multimodal reasoning with Deep Think Benchmarks (developer-reported) SWE-bench Verified 76.2% · GPQA Diamond 91.9% · Terminal-Bench 2.0 54.2% License terms Commercial API Primary source Google, Gemini API deprecations ↗ Last confirmed 26 Sep 2026 Recent change Retired confirmed against Google directly: the models page lists Gemini 3 Pro only under Previous models as (Shut down), and the deprecations table gives gemini-3-pro-preview a shutdown date of 9 Mar… Security & safety Model card and safety evals.
Claude Haiku 4.5	Anthropic	Oct 2025	Text + Vision	Proprietary	Low tier	API	Anthropic model card
Context window 200K tokens Parameters Undisclosed Knowledge cutoff Feb 2025 Lifecycle Superseded Stated use Low-latency, cost-efficient Benchmarks (developer-reported) SWE-bench Verified 73.3% · GPQA Diamond 73.0% · AIME 2025 80.7% (no tools) License terms Commercial API Primary source Anthropic model card ↗ Last confirmed 1 Oct 2026 Recent change Moved to legacy on 7 Oct 2026 when Claude Haiku 5.5 launched; still available on the API. Security & safety System card published; released under the ASL-2 standard.
Grok 2	xAI	Aug 2025	Text	Community	Self-host	Open weights	xAI (Hugging Face)
Context window 128K tokens Parameters Undisclosed Lifecycle Superseded Stated use Open frontier-class weights Benchmarks (developer-reported) HumanEval 88.4% · GPQA 56.0% · MMLU-Pro 75.5% License terms Grok 2 Community License Agreement Primary source xAI (Hugging Face) ↗ Last confirmed 26 Sep 2026 Recent change Confirmed from the cited repo: config.json gives max_position_embeddings 131072 (128K) and the README licenses the weights under the Grok 2 Community License Agreement Security & safety Open weights under the Grok 2 Community License (restricted); deployer owns safety tuning.
GPT-5	OpenAI	Aug 2025	Text + Vision	Proprietary	Mid tier	API	OpenAI Launches ChatGPT Work With GPT-5.6 Agents
Context window 400K tokens Parameters Undisclosed Knowledge cutoff Sep 2024 Lifecycle Deprecated Stated use Broad reasoning and tools Benchmarks (developer-reported) SWE-bench Verified 74.9% · AIME 2025 94.6% (no tools) · MMMU 84.2% License terms Commercial API Primary source OpenAI model card ↗ Last confirmed 12 Oct 2026 Recent change OpenAI deprecated gpt-5-2025-08-07 on 11 Jun 2026 with shutdown on 11 Dec 2026, naming gpt-5.6-sol as the replacement; its model page now calls GPT-5 a previous model. Security & safety System card published.
Qwen3	Alibaba	Apr 2025	Text	Open weights	Self-host	Open weights	Alibaba model card
Context window 128K tokens Parameters 0.6B-235B Lifecycle Superseded Stated use Broad open family, multilingual Benchmarks (developer-reported) LiveCodeBench 70.7% · AIME 2025 81.5% · BFCL v3 70.8% License terms Apache-2.0 Primary source Alibaba model card ↗ Last confirmed 8 Oct 2026 Recent change Alibaba Model Studio no longer lists the April 2025 Qwen3 models, so price is free and access is weights only (checked 8 Oct 2026). Security & safety Apache-2.0 weights; self-host data boundary.
Llama 4	Meta	Apr 2025	Text + Vision	Community	Self-host	Open weights	Meta model card
Context window 1M tokens Parameters up to 400B (MoE) Knowledge cutoff Aug 2024 Lifecycle Current Stated use Open MoE, long context Benchmarks (developer-reported) LiveCodeBench 43.4% · GPQA Diamond 69.8% · MMLU-Pro 80.5% License terms Llama Community License (scale cap) Primary source Meta model card ↗ Last confirmed 26 Sep 2026 Recent change Represents Llama 4 Maverick (1M context, 17B active / 400B total, 128 experts), released 5 Apr 2025. The first-party Meta Model API serves Muse models, not Llama 4, so access is weights-only. Security & safety Open weights: you own the data boundary and the safety tuning. Scale-cap clause for very large deployments.
Gemma 3	Google	Mar 2025	Text + Vision	Community	Self-host	Open weights	Google model card
Context window 128K tokens Parameters 1B-27B Knowledge cutoff Aug 2024 Lifecycle Superseded Stated use Efficient open family Benchmarks (developer-reported) LiveCodeBench 29.7% · GPQA Diamond 42.4% · MMLU-Pro 67.5% License terms Gemma terms of use (restricted) Primary source Google model card ↗ Last confirmed 26 Sep 2026 Recent change Open family (1B/4B/12B/27B); 4B/12B/27B support vision (text plus image). Built on Gemini 2.0 research. ShieldGemma 2 image safety checker built on Gemma 3. Security & safety Open weights under Gemma terms; self-host safety is your responsibility.
DeepSeek-R1	DeepSeek	Jan 2025	Text	Open weights	Self-host	Open weights	DeepSeek model card
Context window 128K tokens Parameters 671B (MoE) Lifecycle Superseded Stated use Open reasoning model Benchmarks (developer-reported) SWE-bench Verified 49.2% · GPQA Diamond 71.5% · MMLU-Pro 84.0% License terms MIT Primary source DeepSeek model card ↗ Last confirmed 8 Oct 2026 Recent change Updated as R1-0528 (May 2025). The DeepSeek pricing page no longer lists R1, so access is weights only (checked 26 Sep 2026). Security & safety MIT weights; self-host data boundary. Reasoning traces can be verbose; review before logging.
Phi-4	Microsoft	Dec 2024	Text	Open weights	Self-host	Open weights	Microsoft model card
Context window 16K tokens Parameters 14B Knowledge cutoff Jun 2024 Lifecycle Current Stated use Small-model reasoning Benchmarks (developer-reported) HumanEval 82.6% · GPQA 56.1% · MMLU 84.8% License terms MIT Primary source Microsoft model card ↗ Last confirmed 26 Sep 2026 Recent change Phi-4 family expanded in 2025 (Phi-4-mini, Phi-4-multimodal, Phi-4-reasoning). Security & safety MIT weights, small enough to run locally; a good fit where data cannot leave the device.
DeepSeek-V3	DeepSeek	Dec 2024	Text	Community	Self-host	Open weights	DeepSeek model card
Context window 128K tokens Parameters 671B (MoE) Lifecycle Superseded Stated use Efficient open MoE, low-cost API Benchmarks (developer-reported) SWE-bench Verified 42.0% · GPQA Diamond 59.1% · MMLU-Pro 75.9% License terms DeepSeek Model License (code is MIT; weights carry use restrictions) Primary source DeepSeek model card ↗ Last confirmed 8 Oct 2026 Recent change License: code MIT, weights under the DeepSeek Model Agreement. The DeepSeek pricing page no longer lists V3, so access is weights only (checked 26 Sep 2026). Security & safety Code MIT; weights under the DeepSeek Model Agreement. Self-host for data control; the first-party DeepSeek API no longer lists V3.
Command R+	Cohere	Aug 2024	Text	Community	Mid tier	API + weights	Cohere model card
Context window 128K tokens Parameters 104B Knowledge cutoff Feb 2023 Lifecycle Superseded Stated use RAG and tool use License terms CC-BY-NC (non-commercial) Primary source Cohere model card ↗ Last confirmed 26 Sep 2026 Recent change Command R+ 08-2024 (current version), a refresh of the original Apr 2024 release; 128K context, CC-BY-NC weights plus paid Cohere API. Superseded by Command A (Mar 2025) and Command A Reasoning. Security & safety Weights non-commercial (CC-BY-NC); enterprise license required to ship commercially.
Mistral Large 2	Mistral	Jul 2024	Text	Community	Self-host	Open weights	Mistral model card
Context window 128K tokens Parameters 123B Lifecycle Retired Stated use Strong European frontier Benchmarks (developer-reported) MMLU 84.0% (pretrained) License terms Mistral Research License (non-commercial); commercial license to self-deploy Primary source Mistral model card ↗ Last confirmed 26 Sep 2026 Recent change The model card marks mistral-large-2407 retired (30 Mar 2025) and names Mistral Large 3 as the replacement; weights remain under the MRL. Security & safety Weights under the Mistral Research License; commercial self-deployment needs a Mistral Commercial License.

About this data

Scope: Model releases the developer itself published specs or pricing for.
Key facts: 41 models from 11 developers, of which 8 ship open weights.
27 of 41 disclose a training-data cutoff; the rest publish none, and this table leaves those cells empty rather than estimating them.
This cycle7 Oct 2026: Anthropic released Claude Haiku 5.5 on October 7, 2026 with a 1M-token context window, 128K-token output and a June 2026 knowledge cutoff, priced from 10 cents per million input tokens and 50 cents per million output tokens. Its row is held as a draft pending a second check. Anthropic now lists Claude Haiku 4.5 as a legacy model, so that row moves from current to superseded. Mistral Large 4, announced on October 6 as a public preview, is now published with 1.05 trillion total and 52 billion active parameters. The Mistral model card lists 1.36 and 4.18 US dollars per million input and output tokens, discounted to 0.68 and 2.09.
Reading the table: A sourced timeline of model releases. Specs and pricing are as published by the developer; “undisclosed” stays undisclosed, and price is a relative tier with the exact figure on the provider’s page via each row’s source. Informational only.

Using this data in research or journalism?

Quoting a figure with a link to this page needs no permission. Reuse of the compiled dataset is licensed under CC BY 4.0: credit SQ Magazine and link back.

Download CSV

What the data shows

The charts below redraw the table above as the trends and comparisons a table cannot show.

Model releases by month

Notable releases per month, dated from the developer's own announcement.

Context windows compared

Tokens of context, in thousands, per the developer's own specification.

Releases by developer

Tracked releases per developer.

Verification ledger

5 most recent of 50 logged updates · RSS

11 Oct 2026 Latest Re-verified (primary source)
Re-confirmed, no change
GPT-5
11 Oct 2026 Updated (primary source)
Bench narrowed to Terminal-Bench 2.1 88.8%
GPT-5.6 Sol
9 Oct 2026 Re-verified
Re-verified and corrected Claude Opus 4.8 and Claude Sonnet 5 against their launch posts and docs pages; both stamped 2026-10-09. GPT-5 and GPT-5.6 Sol remain due.
9 Oct 2026 Updated (primary source)
Security note now reads system card published and cyber safeguards on by default
Claude Sonnet 5
9 Oct 2026 Updated (primary source)
Benchmarks now OSWorld-Verified 83.4% only; GPQA Diamond 93.6% removed as unsourced
Claude Opus 4.8

How this tracker is maintained

Every model passes the same checks before it appears, and the row keeps pace with each new release.

Sourced

Specs, license and pricing come from the developer’s own release: model card, license file, or pricing page. No benchmark screenshots, no third-hand numbers.
Dated

Each row carries its release date and when we last confirmed it. Each version bump gets its own new row.
Re-checked

Reviewed every 30 days because the field moves fast. License changes, price cuts and deprecations are logged in the ledger.

Why not a leaderboard?: Live benchmarks already exist and shift daily. This is the durable record: what shipped, when, under which license, at what cost. Row details carry the scores each developer published at release; for current head-to-head rankings, use a live board such as LMArena.
How often is it updated?: Every 30 days, and immediately when a major model ships. The ledger logs each addition and reclassification.

This is informational content, not a procurement or security recommendation. Specs and pricing are the developer’s own disclosures and change with new versions; benchmark scores are vendor-reported and are not independently reproduced here. Confirm current terms on the developer’s page before acting.