Open AI baseline tracker
Floor / Ceiling
This tracker compares two moving AI baselines: the proprietary ceiling and the open-weight floor.
The ceiling is the best frontier capability available through closed systems such as GPT-5.6 Sol, Terra and Luna, Claude Fable 5, Grok 4.5, Claude Sonnet 5, or Claude Opus 4.8. The floor is the strongest open model you could still use if those proprietary systems disappeared tomorrow.
What the floor means
The floor is not the average open model. It is the strongest openly available fallback for a domain, such as coding, agentic work, or general reasoning.
That matters because the floor tells you what capability has become durable. Once the weights are public and the licence is permissive, the baseline is much harder to remove from the world.
Current snapshot
The July snapshot keeps GLM-5.2, LongCat-2.0, Tencent Hy3, Kimi K2.7-Code, MiniMax M3, Nemotron 3 Ultra, Kimi K2.6, GLM-5.1, DeepSeek V4 Pro, Xiaomi MiMo V2.5 Pro, and Ling-2.6-1T as the current open-floor cohort. GLM-5.2 sets the measured floor because Z.AI released MIT weights and Artificial Analysis scores it across all three domains. LongCat-2.0 and the official Apache-2.0 Tencent Hy3 release are unscored candidates for their exact versions. The proprietary ceiling now splits by domain: GPT-5.6 Sol leads coding and agentic work, while Claude Fable 5 still leads general intelligence.
The points and gaps use Artificial Analysis as of the 10 July 2026 recheck. GPT-5.6 is now generally available, and AA's independent evaluation scores Sol (max) at 59 Intelligence, 80 Coding, and 54 Agentic. Terra scores 55 / 77.4 / 47.4; Luna scores 51 / 74.6 / 45.6. Sol therefore sets the coding and agentic ceilings. Claude Fable 5 remains narrowly ahead on general intelligence at 59.9. Grok 4.5 adds another measured ceiling model at 54 Intelligence, 76 Coding, and 45.7 Agentic.
All three GPT-5.6 tiers have 1.05M-token context windows and 128K maximum output. Official API IDs are gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna; prices are respectively $5/$30, $2.50/$15, and $1/$6 per million input/output tokens. Pro, max, and Ultra are modes rather than extra model IDs. Muse Spark 1.1 is also tracked because Meta's July 9 public preview adds active 1M-context management, MCP and skill use, parallel subagents, computer control, and multimodal perception. Its current evaluation is first-party, so Muse Spark 1.1 does not set an independent gap yet.
Kimi K2.7-Code is the new coding-floor addition. Moonshot released the weights under a modified MIT licence, with a 1T-total / 32B-active MoE architecture, 256K context, image and video input, forced thinking/preserved-thinking behavior, and about 30% lower thinking-token usage than K2.6. I am not moving the displayed AA gap yet because Artificial Analysis has not published a K2.7-Code score.
MiniMax M3 is still tracked as a June floor candidate because MiniMax says M3 is released with 1M context, native multimodality, coding strength, and open-weight plans, and Artificial Analysis gives MiniMax-M3 a 44.4 Intelligence Index score. The caveat is that MiniMax's own page still says the full Hugging Face and GitHub release is coming soon, while Artificial Analysis currently labels the model weights as unavailable, so a strict downloadable-weights reading should treat M3 as a candidate until the public weights land. Nemotron 3 Ultra is tracked because Artificial Analysis gives it a 37.8 Intelligence Index score and lists it as open weights with 550B total parameters, 55B active parameters, a 262K context window, and strong output speed.
LongCat-2.0 is the new open-floor addition. The official model card lists an MIT licence, 1.6T total parameters, roughly 48B active parameters per token, 1M-context training data, LongCat Sparse Attention, N-gram embedding, and integration with Claude Code, OpenClaw, Hermes, and other agent harnesses. Meituan's own benchmark table reports 70.8 on Terminal-Bench 2.1, 59.5 on SWE-bench Pro, 77.3 on SWE-bench Multilingual, 73.2 on FORTE, 79.9 on BrowseComp, 78.8 on RWSearch, 88.9 on GPQA-diamond, and 81.8 on IMO-AnswerBench. Artificial Analysis has a LongCat provider page, but it says it is not currently tracking any LongCat models, so LongCat-2.0 does not move the displayed AA floor yet.
GLM-5.2 remains the measured open floor. Z.AI's model card lists a 744B-total / 40B-active MoE, 1M context, flexible reasoning effort, local serving through SGLang, vLLM, Transformers, and KTransformers, and an MIT open-source licence. Z.AI's GLM-5.2 docs list 1M context and 128K max output, and Z.AI pricing lists $1.40 input, $0.26 cached input, and $4.40 output per million tokens. Artificial Analysis scores GLM-5.2 at 51.1 on Intelligence, 68.8 on Coding, and 43.1 on Agentic work, making it the measured July floor in this tracker. The live provider page lists 15 benchmarked providers; at the July 10 recheck, Blackbox AI was fastest at 447 output tokens per second and DeepInfra (FP4) had the lowest displayed blended price at $0.61 per million tokens. The other open research notes do not move the measured floor: North Mini Code is below the current floor, Gemma 4 12B is a laptop-scale open multimodal model, and DiffusionGemma 26B A4B is an open diffusion LLM focused on high-speed generation.
Sakana Fugu Ultra is a closed orchestration-ceiling candidate, not an open floor model. Sakana describes Fugu and Fugu Ultra as an OpenAI-compatible API where one model dynamically coordinates a pool of expert models; the console lists fugu and fugu-ultra model IDs and says Fugu Ultra can route one to three agents for harder tasks. Sakana's own benchmark table reports Fugu Ultra at 73.7 on SWE Bench Pro, 82.1 on TerminalBench 2.1, 93.2 on LiveCodeBench, and 95.5 on GPQA-D. Sakana's pricing page lists Fugu Ultra at $5 input, $30 output, and $0.50 cached input per million tokens, with higher rates above 272K context. I rechecked Artificial Analysis on 10 July 2026 and found no Sakana or Fugu listing, so it remains a vendor-reported ceiling candidate rather than a gap setter.
Why track it monthly
AI progress is easier to understand when the floor and ceiling are separated. Some months the ceiling jumps. Other months the floor catches up. The important long-run question is not whether one lab is briefly ahead, but how much capability remains available even without the frontier APIs.