GPT-5.6 Sol
The flagship top-tier model in the GPT-5.6 series. It achieves a score of 53.6 on the Agent Ultimate Exam and 80 on the Coding Agent Index. It supports an Ultra mode enabling four agents to run in parallel.
Core Facts
Timeline (last 90 days)
In one test case, GPT-5.6 Sol cost 32.5 times Jev, and Opus 5 cost 184.9 times Jev.
In its tests, SerpApi used GPT-6.1 Sol (low reasoning effort) as the answer model, with scoring performed by GPT-5.6 Sol (high reasoning effort) without access to search tools.
V4.1 Flash 1017 costs about $0.20 per task, while GPT-5.6 Sol costs about $3.20.
Under the reference workload (1,100 tasks per month, each with 12K fresh input, 48K cached input, and 6K output tokens), enabling caching saves about $190 per month compared with not enabling it.
The DeepSeek comparison table shows that the new model beats Kimi K3 on 10 of 11 tests, beats GPT-5.6 Sol on 9, but trails Claude Opus 5 on 7.
Artificial Analysis previously measured the V4-Flash series and found its per-task cost far lower than competitors (3 cents in its testing framework, versus $1.86 for GPT-5.6 Sol).
Mustafa Suleiman said MAI Base 2 is stronger than the model powering GPT-5.6 Sol and used only a quarter of the compute OpenAI used to train that model, with this comparison based on non-public information Microsoft obtained as an investor.
The OpenRouter model pages show that across all traffic, DeepSeek V4.1 Flash accounts for about 36% of tokens and GPT-6.1 Sol about 18%.
Claude's vocabulary distance from humans dropped from 0.064 in Opus 5 to 0.052 in Opus 5.5, while OpenAI rose from 0.101 in GPT-5.6 Sol to 0.109 in GPT-6 Astra.
Security startup XBOW used OpenAI's GPT-5.6 Sol to chain three Chrome zero-days and successfully escape the browser sandbox on Windows.
40 more timeline events
All Facts (20)
GPT-5.6 Sol is an all-new flagship model focused on complex reasoning and agentic programming.
75%VerifiedDevin officially announced that the newly integrated GPT-5.6 Sol model is now live on both the Devin Desktop and Devin CLI platforms
65%UnverifiedA developer characterized GPT-5.6 Sol as the most 'sleepy' model he had used and described its behavior as 'laziness'
80%UnverifiedExcessive model caution creates misalignment between internal safety logic and users' expectations for productivity, which is especially costly in paid API scenarios where users pay for waiting time
75%UnverifiedThe release party claimed that the cost of DeepSeek-V4.1-Flash is only 1/40 that of Opus 5 and GPT-5.6 Sol.
70%UnverifiedA developer reported that GPT-5.6 Sol spends approximately 70% of its runtime executing sleep commands
70%UnverifiedIn Codex, a 1 million token context window can be unlocked for GPT-5.6 Sol by adding the model slug and manually setting the context window to 1 million and the auto-compact threshold to 900,000.
60%UnverifiedIn one test case, GPT-5.6 Sol cost 32.5 times Jev, and Opus 5 cost 184.9 times Jev.
50%UnverifiedIn its tests, SerpApi used GPT-6.1 Sol (low reasoning effort) as the answer model, with scoring performed by GPT-5.6 Sol (high reasoning effort) without access to search tools.
50%UnverifiedV4.1 Flash 1017 costs about $0.20 per task, while GPT-5.6 Sol costs about $3.20.
50%UnverifiedUnder the reference workload (1,100 tasks per month, each with 12K fresh input, 48K cached input, and 6K output tokens), enabling caching saves about $190 per month compared with not enabling it.
50%UnverifiedThe DeepSeek comparison table shows that the new model beats Kimi K3 on 10 of 11 tests, beats GPT-5.6 Sol on 9, but trails Claude Opus 5 on 7.
50%UnverifiedArtificial Analysis previously measured the V4-Flash series and found its per-task cost far lower than competitors (3 cents in its testing framework, versus $1.86 for GPT-5.6 Sol).
50%UnverifiedMustafa Suleiman said MAI Base 2 is stronger than the model powering GPT-5.6 Sol and used only a quarter of the compute OpenAI used to train that model, with this comparison based on non-public information Microsoft obtained as an investor.
50%UnverifiedThe OpenRouter model pages show that across all traffic, DeepSeek V4.1 Flash accounts for about 36% of tokens and GPT-6.1 Sol about 18%.
50%UnverifiedClaude's vocabulary distance from humans dropped from 0.064 in Opus 5 to 0.052 in Opus 5.5, while OpenAI rose from 0.101 in GPT-5.6 Sol to 0.109 in GPT-6 Astra.
50%UnverifiedFour researchers spent six days during Black Hat and DEF CON, with a token bill of about $40,000; the AI agent was responsible for searching code and drafting exploit stages, while humans chose targets and provided context.
50%UnverifiedThe evaluation uses two AI judges: Anthropic's Claude Opus 5.5 and OpenAI's GPT-5.6 Sol, who cannot see each other's judgments, with bias eliminated through anonymizing and randomly ordering answers, a second round of letter-reordering review, and setting up control tests.
50%UnverifiedOn TerminalBench, HydraFusion has the same cost as GPT-5.6 Sol but achieves 11.4 points higher quality.
50%UnverifiedSol's factual error rate is approximately half that of GPT-5.6 Sol.
50%