Changelog
Updates to the AI Frontier Model Tracker - new models, data corrections, benchmark additions, and feature changes.
August 27, 2026
GLM-5.3-Flash and Qwen3.8-Flash-Next, and an Ox Alpha correction
- Correction. On August 23 we said the anonymous ox-alpha endpoint was probably GLM-5.3 under a codename, and therefore already tracked. That was wrong. Z.ai confirmed at launch that ox-alpha was GLM-5.3-Flash, a distinct model that was not in the tracker. The tokenizer probes matching 30-for-30 correctly identified the model family; reading that as the same model was the error. Both the August 23 entry and this one now say so.
- Added GLM-5.3-Flash (Zhipu AI, August 26). Despite the name it is not a distillation of GLM-5.3 - 320B MoE with 18B active against GLM-5.3's 743B/40B, trained on its own 30T-token multimodal corpus, and the first natively multimodal model in the GLM-5 family. Linear attention for local dependencies, sparse attention for global. 1M context, image and video in. MIT weights on Hugging Face at release, which GLM-5.3 still does not have. Z.ai reports Terminal-Bench 2.1 84.3 against Claude Opus 4.8's 85.0, DeepSWE v1.1 63.4 from GLM-5.2's 46.2, AutomationBench 48.8 from 26.2, and OfficeQA Pro 62.4 ahead of Opus 4.8. AA Intelligence Index v4.1.1 57. $0.15/$0.50 per 1M, halved through September 9.
- Added Qwen3.8-Flash-Next (Alibaba/Qwen, August 26). A deliberate preview of the Qwen4 architecture, shipped ahead of the full family so tooling and quantization can adapt. 125B MoE with 6B active across 512 experts, plus a 51B n-gram embedding table and a 4B multi-token prediction layer, 176B total on disk. 262K native context extensible to 1M, multimodal in, open weights on ModelScope in standard and FP8. Vendor-reported GPQA Diamond 91.7, SWE-bench Pro 62.5, DeepSWE v1.1 58.7, CoWorkBench 73.9. Runs locally in roughly 75GB of RAM. Included for the architecture and the GPQA figure rather than the parameter tier.
- GLM-5.3 weights still pending. The zai-org/GLM-5.3 repository remains a placeholder listing August 28. The entry stays Closed and now carries that date in its description; flip it to Open and add the Hugging Face repo when the weights land.
- 100 models across 20 providers
August 23, 2026
Grok 4.6, Gemini 3.7 Flash, GLM-5.3
- Added Grok 4.6 (xAI, August 12). Flagship refresh barely a month after 4.5, aimed at long-running agents and agentic coding. AA Intelligence Index 61 - level with GPT-5.6 Sol, one point behind Claude Fable 5. The gains sit in agentic coding: DeepSWE v1.1 65.9 from 54.0, APEX-Agents 57.5 from 47.1. Also published: CursorBench v3.2 69.9%, FrontierCode v1.1 61.3%, APEX-SWE 56.4%, Terminal-Bench v3.0 26%, GDPVal-AA v2 1753, AA-Briefcase 1577, Harvey LAB (Vals) 15.8%. 500K context, cutoff February 1, 2026. $2/$6 per 1M, fast variant at double. Standard benchmark columns left blank - xAI published none of them.
- Added Gemini 3.7 Flash (Google, August 13), 23 days after 3.6 Flash. 1,048,576-token input, 65,536 output. Multimodal in (text, image, video, audio, PDF), text out. Google reports DeepSWE v1.1 65.3 from 49.0, FrontierCode 1.1 Main 43.6 from 34.4, AutomationBench 30.4 from 17.0. AA Intelligence Index 56 against 52 for 3.6 Flash, and first of 186 models on output speed at 340.1 tokens/sec. Pricing has an expiry: $0.75/$3.75 is introductory through December 31, 2026, then doubles to $1.50/$7.50 on January 1 - exactly what 3.6 Flash costs today. The tracker records the current rate; it will need revisiting in January.
- Added GLM-5.3 (Zhipu AI, August 14). Same 743B MoE / ~40B active base as GLM-5.2 - every gain comes from extended post-training rather than a new pre-train, which is the interesting part. 1M context, 128K output. Z.ai reports Terminal-Bench 3.0 28.3 from 4.6, DeepSWE v1.1 66.9 from 46.2, Agents' Last Exam 28.5 from 23.8, and a 50% coding gain on its in-house Code Bench. The larger jump is cybersecurity: CyberGym 84.5% from 77.2%, ExploitBench 54.4% from 24.4%. Listed as Closed because the weights were not public at time of entry - Z.ai committed to publishing to zai-org on Hugging Face roughly two weeks after launch, once safety evaluation finishes. Flip to Open when they land. Launch was widely reported August 14; z.ai's own release notes log it under August 18.
- Considered and left out. GPT-5.6-Cyber (OpenAI, August 10) - specialised variant rather than a frontier release. DeepSeek-V4-Pro-0813 (August 13) - a checkpoint refresh of the V4 Pro entry already tracked from April 24, and its claimed gains have not been replicated by any third party. Qwen3.8-27B (August 14) - 27B open model, well below the bar set by Qwen3.8-Max. GLM-5.2 Turbo - derivative speed variant.
- Ox Alpha, the stealth model that appeared anonymously on OpenRouter on August 20, is not listed. No lab has claimed it, and tokenizer probes matching GLM-5.3 30-for-30 plus a stack trace exposing Zhipu internal API classes suggest it is GLM-5.3 under a codename - in which case it is already here under its real name. Revisit if a lab claims it as a distinct model.
Corrected August 27: wrong. Z.ai confirmed at launch that ox-alpha was GLM-5.3-Flash, a separate model that was not in the tracker at all. The tokenizer evidence was sound - it identified the family - but the conclusion drawn from it was not. Added below. - 98 models across 20 providers
August 11, 2026
Nemotron 3.5 Lightning (NVIDIA) - 20th provider
- Added Nemotron 3.5 Lightning (NVIDIA, August 11), and NVIDIA as the 20th provider. 30B Mixture-of-Experts with just 3B active per token on a Mamba2-Transformer hybrid with Multi-Token Prediction - the low active count is the design goal, buying up to 4x the output speed of similar-sized models. Open weights under OpenMDW-1.1 in BF16 and NVFP4. Context up to 1M, though NVIDIA runs 256K for single-H100 deployment. Pre-trained on 20T+ tokens across English, 19 other spoken languages, and 43 programming languages. Reasoning toggles on and off via the chat template. Text only.
- Model-card benchmarks (primary source, not vendor marketing): MMLU-Pro 81.94, GPQA Diamond 75.44, SWE-bench Verified 51.56, IFBench (loose) 71.88, AA-LCR 52.00, Terminal-Bench 2.1 24.58. NVIDIA separately reports 86% on its own PinchBench, completing 10,000 tasks 30% faster than Qwen3.6 35B.
- Worth reading against Muse Glimmer, added a day earlier: both 30B open-weight models sized for a single GPU, pulling in opposite directions. Glimmer takes the agentic and long-context benchmarks, Lightning takes raw knowledge and speed - MMLU-Pro 81.94 and GPQA 75.44 against a Terminal-Bench 2.1 of 24.58. NVIDIA is explicit that this one targets high-volume, low-latency agent execution rather than frontier reasoning.
- 95 models across 20 providers
August 11, 2026
Catch-up: Gemini 3.6 Flash, two Meta models, DeepSeek V4 Flash GA
- Added Gemini 3.6 Flash (Google, July 21). 1,048,576-token input, 65,536 output. $1.50/$7.50 per 1M, cached input $0.15 - a lower output rate than 3.5 Flash. Multimodal in (text, image, video, audio, PDF), text out. Google reports ~17% fewer output tokens than 3.5 Flash while beating it on every published benchmark: SWE-Bench Pro 58.7 vs 55.1, OSWorld-Verified 83.0 vs 78.4. Artificial Analysis Intelligence Index 52 at high effort. Standard benchmark columns left blank - Google published none of them for this release.
- Added Muse Spark 1.2 (Meta, August 5). Flagship coding model behind Muse Code, a terminal agent for macOS and Linux. 1M context, unchanged from 1.1. Mandatory reasoning across five effort levels. $1.25/$4.25 per 1M standard, or $0.10/$0.20 on a contributor tier that opts into Meta training at 60 req/min. Vendor-run only: Terminal-Bench 2.1 82.9%, DeepSWE v1.1 59.3%, AA Intelligence Index 54 (tied with Grok 4.5). Weights promised, no date or license.
- Added Muse Glimmer (Meta, August 10). ~30B dense, Apache 2.0, sized to run on one consumer GPU - under 20GB quantized, fits 24GB/32GB. Text and image in, text out. 131K context, cutoff January 4, 2026. Meta-published: AIME 2026 94.7, MCP Atlas 75.5, SWE-Bench Pro 51.2, DeepSearch QA 74.6, IFBench 77.0, OSWorld-Verified 65.9. Only AIME maps to a tracker column; the rest sit in the description. Note Meta's comparison table picks the better of each competitor's self-reported or Meta-reproduced score.
- Updated DeepSeek V4 Flash from the April 24 preview to the July 31 GA build (DeepSeek-V4-Flash-0731). Not a new architecture - the April checkpoint retrained through a reworked post-training pipeline for coding, agents, reasoning, and tool use. DeepSeek reports it beats the larger V4-Pro preview on every agentic benchmark it published. MIT weights supersede the preview; API in public beta. Benchmark columns still reflect the April preview.
- Fix: a null price rendered as "free" regardless of access type, so seven closed models with undisclosed rates - Qwen3.8-Max, Muse Spark, GPT-Rosalind, Qwen3.6-Max-Preview, GPT-5.5 Instant, MAI-Code-1-Flash, and MAI-Thinking-1 - were shown as free to call. Open-weight models still show "free" (self-hosted); closed models with no published rate now show a dash.
- 94 models across 19 providers
August 5, 2026
Qwen3.8-Max goes GA (Alibaba)
- Updated Qwen3.8-Max from preview to general availability (Alibaba, August 3). Now callable via API (QwenCloud) with open weights due the following week - Alibaba's first open-weight model at Max scale. Active-parameter count now disclosed: 2.4T total with ~95B active per token; 1M context confirmed. Vendor-reported benchmarks (not independently verified): PaperBench 93.0 vs Fable 5's 88.8, IFBench 82.8, Terminal-Bench 2.1 86.6 (ahead of Opus 4.8 and Fable 5, behind GPT-5.6 Sol). Numeric benchmark fields left blank pending independent verification.
- Watch: Gemini 3.5 Pro still unreleased - no model card or official pricing.
- 91 models across 19 providers
July 25, 2026
Claude Opus 5 (Anthropic)
- Added Claude Opus 5 (Anthropic, July 24) - a new flagship positioned as a daily-driver Opus that rivals Fable 5 at half the price ($5/$25 vs $10/$50). Anthropic reports it more than doubles Opus 4.8 on Frontier-Bench v0.1, comes within 0.5% of Fable 5 on CursorBench 3.2 at half the cost, and roughly 3x the next-best model on ARC-AGI 3. Standard SWE-bench / GPQA numbers not published at launch, so those fields are left blank. 1M context, multimodal, adaptive thinking.
- 91 models across 19 providers
July 20, 2026
Qwen3.8-Max-Preview (Alibaba)
- Added Qwen3.8-Max-Preview (Alibaba, July 19) - the flagship Qwen3.8, at 2.4T parameters Qwen's first multimodal model above 1T, callable now via Alibaba's Token Plan / Qoder. Open weights promised "soon" (no date, no Hugging Face card yet); active-parameter count, context window, and per-token pricing undisclosed. Qwen positions it as "second only to Fable 5" on internal evaluations - no independent benchmarks published yet, so those fields are left blank.
- Watch: Gemini 3.5 Pro still unreleased - no model card or official pricing.
- 90 models across 19 providers
July 18, 2026
Kimi K3 (Moonshot AI)
- Added Kimi K3 (Moonshot AI, July 16) - a 2.8T-parameter open-weight multimodal reasoning MoE (896 experts, 16 active per token) built on Kimi Delta Attention, with 1M context. Independent testing ranks it #4 among all frontier models, behind only Claude Fable 5 and GPT-5.6 Sol and ahead of Opus 4.8; debuted #1 on LMArena's Frontend Code Arena (1679 Elo). $3/$15 per 1M tokens; full open weights due July 27.
- Watch: Gemini 3.5 Pro still unreleased - missed its July 17 target (a third slip), with no model card or official pricing yet.
- 89 models across 19 providers
July 16, 2026
Inkling (Thinking Machines Lab)
- Added Inkling (Thinking Machines Lab, July 15) - the lab's debut model and a new provider (Mira Murati's team). Sparse MoE with 975B total / 41B active parameters, pretrained on 45T tokens of text, image, audio, and video. 1M context, open weights on Hugging Face, fine-tunable via Tinker. The lab is candid that it is "not the strongest overall model available today," emphasizing multimodal breadth and efficient reasoning. Self-reported at effort=0.99 (not independently verified): GPQA Diamond 87.2, AIME 2026 97.1, SWE-bench Verified 77.6. No first-party per-token pricing disclosed yet.
- Watch: Gemini 3.5 Pro still in preview, targeting July 17.
- 88 models across 19 providers
July 10, 2026
Grok 4.5, GPT-5.6 family (Sol, Terra, Luna)
- Added Grok 4.5 (xAI, July 8) - new flagship positioned as Opus-class, built on the 1.5T V9 foundation and trained on real Cursor session data for coding and agentic work. Artificial Analysis Intelligence Index 54 at launch (#4 overall, up 16 points over Grok 4.3); xAI claims ~4.2x fewer output tokens than Opus 4.8 on SWE-Bench Pro. 500K context, $2/$6 per 1M.
- Added the GPT-5.6 family (OpenAI, public July 9 after a gated preview from June 25) as three tiers, matching the GPT-5.5 pattern: Sol (flagship, SOTA on Terminal-Bench 2.1, $5/$30), Terra (mid, GPT-5.5-class at ~half the cost, $2.50/$15), and Luna (fastest and cheapest, $1/$6). All 1M context, with programmatic tool calling in the Responses API.
- Watch: Gemini 3.5 Pro still in preview, targeting July 17.
- 87 models across 18 providers
July 8, 2026
Claude Sonnet 5; Fable 5 & Mythos 5 back online
- Added Claude Sonnet 5 (June 30) - Anthropic's most agentic Sonnet, positioned close to Opus 4.8 at a lower price; replaced Sonnet 4.6 as the default on Claude.ai (Free/Pro). Humanity's Last Exam 34.6% (46.8% with tools), OSWorld-Verified 78.5%. 1M context, newer tokenizer (~30% more tokens). Intro pricing $2/$10 per 1M through Aug 31, then $3/$15 (batch 50% off). SWE-bench Verified and GPQA numbers vary across sources, pending system-card confirmation.
- Claude Fable 5 and Mythos 5 redeployed (July 1) - the US lifted the June 12 export-control order, so both are available again across all platforms. Updated their entries.
- Added LongCat-2.0 (Meituan, June 29) - a 1.6T open MoE agentic coding model (~48B active, 1M context) trained on a ~50,000-card domestic Chinese ASIC cluster. Ran anonymously as "Owl Alpha" and topped OpenRouter's coding usage charts for ~2 months before Meituan revealed it - added on demonstrated adoption. Vendor-reported benchmarks, independent verification pending: 59.5 SWE-Bench Pro, 70.8 Terminal-Bench 2.1, 77.3 SWE-Bench Multilingual. MIT license, full weights coming (INT8 posted). Meituan is the 18th provider.
- Watch (not added): GPT-5.6 (Sol/Terra/Luna) previewed June 26 but gated to ~20 US-gov-approved orgs; Gemini 3.5 Pro delayed to July 17 (architecture rebuild); Grok 4.5 in private beta.
- 83 models across 18 providers
June 27, 2026
Doubao Seed 2.1 Pro & Turbo (ByteDance)
- Added Doubao Seed 2.1 Pro (June 24) - ByteDance's first entry in the tracker. Flagship deep-thinking model for the coding/agent era, which ByteDance positions as comparable to GPT-5.5. Leads GDPval, MobileWorld, CharXiv-RQ, MeasureBench; #8 on Code Arena Frontend (1539, level with Claude Opus 4.6); top-tier on Agents' Last Exam. Multimodal (image + video). Closed, API via Doubao / Volcano Engine. ¥6/¥30 per 1M (~$0.83/$4.17). Parameters and context window not disclosed.
- Added Doubao Seed 2.1 Turbo (June 24) - low-cost, low-latency sibling for high-volume enterprise deployment, reported comparable to Pro. ¥3/¥15 per 1M (~$0.42/$2.08).
- Watch: Gemini 3.5 Pro (2M context, Deep Think) still in limited Vertex preview - no GA, no published specs/pricing/benchmarks, so not added.
- 81 models across 17 providers
June 21, 2026
Kimi K2.7 Code, GLM-5.2
- Added Kimi K2.7 Code (June 12) - Moonshot AI's coding-specialized 1T MoE / 32B active, built on K2.6. Mandatory thinking mode, ~30% fewer reasoning tokens than K2.6; MoonViT vision encoder for image and video input; native INT4. Moonshot reports +21.8% on Kimi Code Bench v2 over K2.6 - no independent SWE-Bench numbers published yet. 256K context. Open-weight, Modified MIT. $0.95/$4.00.
- Added GLM-5.2 (June 13) - Zhipu AI open-weight 753B MoE / 40B active. SWE-Bench Pro 62.1% (up from GLM-5.1's 58.4%), beating GPT-5.5 (58.6%). Terminal-Bench 2.1 81.0%, AIME 2026 99.2%, GPQA Diamond 91.2%, HLE 54.7% with tools. Selectable reasoning effort (High/Max). 1M context, up to 131K output. MIT license. $1.40/$4.40; GLM Coding Plan from ~$10/mo.
- Noted June 12 export-control suspension of Claude Fable 5 and Mythos 5 - Anthropic pulled both across all platforms (Anthropic API, AWS Bedrock, Google Cloud, Microsoft Foundry, Snowflake, Box) after a US directive it could not enforce by nationality in real time.
- 79 models across 16 providers
June 11, 2026
Claude Mythos 5, Claude Fable 5
- Added Claude Mythos 5 (June 9) - Anthropic's restricted Mythos-class frontier. Per System Card: 95.5% SWE-Bench Verified, 80.3% SWE-Bench Pro, 88.0% Terminal-Bench 2.1, 85.0% OSWorld-Verified, 94.1% GPQA Diamond (Anthropic now considers GPQA saturated), 59.0% HLE (64.5% w/tools), 88.0/93.3% BrowseComp single/multi-agent. 1M context. $10/$50. Restricted to Project Glasswing partners (US gov cyberdefenders) and select biomedical research orgs.
- Added Claude Fable 5 (June 9) - same underlying model as Mythos 5, with safety classifiers that fall back to Opus 4.8 for cyber/bio/chem/distillation queries. Per System Card: 95% SWE-Verified, 80% SWE-Pro, 84.3% Terminal-Bench 2.1, 85.0% OSWorld, 29.3 FrontierCode Diamond, 1932 GDPval-AA, 57.9% OfficeQA Pro. Anthropic cites Stripe's 50M-line Ruby migration in one day. 1M context. $10/$50 (batch $5/$25). Pro/Max/Team include Fable 5 free through 2026-06-22. Available via Anthropic API, AWS Bedrock, GitHub Copilot.
- Updated MiniMax M3 desc - weights commitment of ~June 11 has now slipped; huggingface.co/MiniMaxAI still showed only M2-series repos as of today.
- 77 models across 16 providers
June 4, 2026
MAI-Code-1-Flash, MAI-Thinking-1, MiniMax M3
- Added MAI-Code-1-Flash (June 2) - Microsoft's first in-house coding model. 5B active sparse MoE (total parameter count not published on Microsoft's pages). 51.2% SWE-Bench Pro (+16 over Claude Haiku 4.5). +28.9 IF Bench, 85.8% adversarial accuracy. Solves harder problems with up to 60% fewer tokens on SWE-Verified. Rolling out to VS Code GitHub Copilot.
- Added MAI-Thinking-1 (June 2) - Microsoft's first in-house reasoning model, not a distillation. 35B active sparse MoE (total not published). Per Build 2026 keynote: 52.8% SWE-Bench Pro (alongside Opus 4.6), 97.0% AIME 2025, 94.5% AIME 2026. 256K context. Private preview on Microsoft Foundry.
- Added MiniMax M3 (June 1) - open-weights-committed model combining frontier coding, 1M context, and native multimodality (image + video). Weights and technical report pledged for ~June 11. Built on MiniMax Sparse Attention (MSA) - 9x faster prefill, 15x faster decoding at 1M context. 59.0% SWE-Bench Pro vs Opus 4.8's 69.2% (released four days earlier), 66.0% Terminal-Bench 2.1, 70.06% OSWorld-Verified, 74.2% MCP Atlas, 83.5 BrowseComp. $0.60/$2.40 standard ($0.30/$1.20 promotional at launch).
- 75 models across 16 providers
May 28, 2026
Claude Opus 4.8, Grok Build 0.1, Cohere Command A+, Qwen3.7-Max GA
- Added Claude Opus 4.8 (May 28) - 88.6% SWE-Verified, 69.2% SWE-Pro, 93.6% GPQA Diamond, 83.4% OSWorld-Verified, 84% Online-Mind2Web. 1M context, $5/$25 standard, new Fast Mode at $10/$50 (3x cheaper than 4.7 Fast Mode). All benchmarks verified verbatim against the official Opus 4.8 System Card PDF.
- Added Grok Build 0.1 (May 20) - first xAI model purpose-built for agentic coding. 256K context, text+image, $1/$2 per 1M tokens. 88.9% PinchBench.
- Added Cohere Command A+ (May 20) - 218B sparse MoE / 25B active, Apache 2.0, multimodal (text+image+tool use). 128K input / 64K output. 75.1% MMMU, 80.6% MathVista, 37 Intelligence Index. Runs on two H100s at W4A4.
- Updated Qwen3.7-Max (May 20) - graduated from preview at Alibaba Cloud Summit. 92.4% GPQA Diamond, $2.50/$7.50, cached input $0.25. Lowest hallucination rate among frontier models (22.9%).
- 72 models across 16 providers
April 25, 2026
GPT-5.5, DeepSeek V4, Tencent Hy3
- Added GPT-5.5 (April 23) - first fully retrained base since GPT-4.5, 1M context, multimodal, $5/$30
- Added GPT-5.5 Pro (April 24) - higher-accuracy variant, $30/$180
- Added DeepSeek V4 Pro (April 24) - 1.6T/49B active, 90.1% GPQA Diamond, 80.6% SWE-bench, MIT license
- Added DeepSeek V4 Flash (April 24) - 284B/13B active, 88.1% GPQA Diamond, $0.14/$0.28, MIT license
- Added Tencent Hy3 Preview (April 23) - 295B/21B active, 87.2% GPQA Diamond, 74.4% SWE-bench, new provider
- DeepSeek V4 and Hy3 benchmarks verified against HuggingFace model cards
- GPT-5.5 benchmarks nulled pending primary source verification (OpenAI blog inaccessible)
- 66 models across 16 providers
April 23, 2026
7 new models, MCP server improvements
- Added Claude Opus 4.7 (April 16) - 87.6% SWE-bench Verified, 94.2% GPQA Diamond, xhigh effort level, 3.75MP vision
- Added GPT-Rosalind (April 16) - first domain-specific OpenAI model for life sciences
- Added Grok 4.3 Beta (April 17) - 2M context, native video understanding, SuperGrok Heavy only ($300/mo)
- Added Qwen3.6-Max-Preview (April 20) - top scores on 6 coding benchmarks, free preview on Bailian
- Added Qwen3.6-27B (April 22) - dense 27B open-weight vision model, 87.8% GPQA Diamond, 77.2% SWE-bench, 83.9% LiveCodeBench v6, Apache 2.0
- Added GPT-5.3-Codex (Feb 24) - agentic coding model, SOTA on SWE-Bench Pro, 400K context, $1.75/$14
- Added Gemini 3.1 Flash Lite (March 3) - most cost-efficient Google model, 1M context, $0.25/$1.50 per 1M tokens
- MCP server: improved tool descriptions with attribution URLs, reduced default limit from 50 to 25, disambiguated get/search tool pairs, added freshness cadence info
- 60 models across 15 providers
April 17, 2026
Qwen3.6-Plus and Qwen3.6 35B-A3B
- Added Qwen3.6-Plus (April 2, 2026) - proprietary flagship with 1M native context, 78.8% SWE-bench Verified
- Added Qwen3.6 35B-A3B (April 15, 2026) - open-weight MoE with 3B active / 35B total params, runs on laptops, Apache 2.0
- 44 models across 15 providers
April 12, 2026
Meta Muse Spark, Cohere Command A, nav restructure, production prep
- Added Meta Muse Spark (April 8, 2026) - first model from Meta Superintelligence Labs
- Added Cohere Command A (March 13, 2025) - 111B enterprise RAG model
- All provider logos now local (Anthropic, Google, Meta, Qwen, AWS, Cohere, IBM)
- Download weight icons from LobeHub (HuggingFace, Ollama dark mode)
- Resources nav restructured as mega menu with Research & Data section
- /research/ landing page created
- Newsletter forms fixed (Pipedrive marketing_status, correct worker URL)
- Subscribers differentiated: General Newsletter vs AI Model Tracker Updates
- Exact release dates for all 42 models (verified against official sources)
- Full site audit: 407 pages, 100% schema coverage, AEO avg 72
- 42 models across 14 providers
April 11, 2026
Selectable benchmark columns, new subpages, structural overhaul
- 10 benchmark columns now available (MMLU-Pro, GPQA Diamond, SWE-bench, HumanEval, LiveCodeBench, MATH, AIME, HLE, $/M In, $/M Out)
- Column picker - choose up to 5 visible columns, persists in localStorage
- Default sort changed to GPQA Diamond
- Cost calculator page
- JSON API endpoint (CC BY-NC 4.0)
- Releases timeline page with RSS feed
- Newsletter signup for tracker updates
- Sticky filter bar and table headers
- Hero redesigned with card links to subpages
- All URLs moved to /research/demandsphere-radar/ai-frontier-model-tracker/
- Proper breadcrumbs on all pages
- Added to nav menu and resource center
April 10, 2026
Data corrections and new models
- Full data audit - corrected 30+ benchmark scores, 6 release dates, 3 context windows, 4 pricing values
- Added Gemma 4 31B, Gemma 4 26B-A4B, Qwen3.5 397B-A17B, GPT-5.4 Mini
- Removed legacy models (GPT-4o, o1, GPT-5.2, Claude 3.5 Sonnet)
- Citations tab added with per-model source links
- Recent Releases feed section
- Provider count now dynamic
- Fixed license claims (Kimi K2 Apache->Modified MIT, Hermes 405B Apache->Llama 3.1)
- Set unpublished scores to null (Llama 4 HumanEval/MATH, DeepSeek HumanEval)
April 9, 2026
Initial launch
- 40 frontier models across 14 providers
- Sortable table with MMLU-Pro, HumanEval, MATH benchmarks
- Expandable rows with Detail, Benchmarks chart, News feed, Download Weights
- Filters: type, access, SOTA, multimodal
- Provider logos for OpenAI, xAI, DeepSeek, Moonshot, MiniMax, Mistral, Microsoft, Nous
- FAQ schema, OG image