Concluded·following since 6 Aug 2026·Day 7·8 sources·updated 14 Aug 2026

OpenAI GPT-5.6 Sol and Luna

The gist

OpenAI unified GPT-5.6 Sol across ChatGPT tiers and previewed a 14x-speed Ultrafast API mode via Cerebras

The GPT-5.6 family lets builders achieve results comparable to frontier models at substantially lower cost by mixing model sizes and reasoning effort levels. The Maxwell Conjecture disproof is a documented case of AI contributing a key construction idea in original mathematical research.

The full picture

GPT-5.6 Sol is OpenAI's unified model for both Instant and deep reasoning modes across ChatGPT Plus and Pro, with a reasoning effort slider replacing what had felt like switching between separate models. Free and Go users received GPT-5.6 Luna as their default, with unlimited text chats and a Think button. OpenAI's internal evaluations in finance, medicine, and law found factual errors were 68% less common with Sol compared to GPT-5.5 Instant. On August 13, OpenAI previewed Ultrafast mode, an API service tier running GPT-5.6 Sol at up to 14 times normal speed via a Cerebras partnership, delivering up to 750 output tokens per second. A builder's guide published the same day showed that enabling retained reasoning across turns and native compaction raised Sol's ARC-AGI-3 public-set score from 13.3% to 38.3% using roughly 6x fewer output tokens, with no model changes. Luna at Extra High reasoning matched GPT-5.5 Extra High on BrowseComp (84.04% vs 84.36%) at $1.33 versus $33.27. Three API primitives shipped alongside: persistent reasoning across turns, native multi-agent orchestration, and programmatic tool calling that processes tool outputs outside the model context window. Prompt cache TTL was also extended to a minimum of 30 minutes with deterministic cache breakpoints. Separately, mathematicians Arathoon, Ball, and Kvalheim disproved the Maxwell Conjecture, a problem from the 1870s, by constructing a five-charge counterexample with 24 critical points; they stated they were grateful to GPT-5.6 Sol for suggesting the key construction.

How it developed

14 August 2026

OpenAI published a builder's guide on August 13 showing Luna at Extra High matching GPT-5.5 Extra High on BrowseComp at one-twenty-fifth the cost, and announced Ultrafast mode for Sol, up to 14x faster, with API access limited to select customers.

The same day, DeepSeek-V4-Pro launched with native OpenAI Responses API compatibility and scored within 15 Code Arena WebDev points of Sol Extra High at roughly one-thirty-first the blended price, while Grok 4.6 scored level with Sol on Artificial Analysis's overall benchmark at about half the turns and input tokens for extended agentic tasks.

First citedRohan Paul Twitter The Neuron DeepSeek Twitter Miles Brundage Twitter+1

13 August 2026

Builder's guide published with ARC-AGI-3 details, BrowseComp cost comparison, three API primitives, and extended cache TTL

11 August 2026

Model ML published results on August 10 from its Composite benchmark comparing GPT-5.6 Sol, OpenAI's unified reasoning model for ChatGPT Plus and Pro, against Anthropic's Opus 5 and Fable 5 in finance document automation.

Sol used 21% fewer tokens per PowerPoint deck than Fable 5 and 36% fewer tokens per Excel workbook than Opus 5, completed all PowerPoint workflow test cases versus 76% for Opus 5, and cut tearsheet build time at one global asset manager from roughly one hour to about five minutes.

First citedThe Neuron Together AI Blog OpenAI Blog

10 August 2026

Model ML publishes case study showing Sol outperforms Opus 5 and Fable 5 on finance document automation benchmarks

8 August 2026

Altman argues inference scale at modest margins is sufficient to fund frontier model training

OpenAI on August 8 cut Luna API prices 80% to $0.20 input and $1.20 output per million tokens, cut Terra prices 20%, and announced Fast Mode for Sol on the API; it also disclosed that changing two API settings raised GPT-5.6 Sol's ARC-AGI-3 public-set score from 13.3% to 38.3%. Both follow an August 7 announcement that GPT-5.6 Sol now powers both Instant and Thinking modes for Plus and Pro users via a reasoning slider, with internal evaluations finding factual errors 68% less common versus GPT-5.5 Instant.

First citedZvi's AI Roundups Rohan Paul Twitter The Neuron

7 August 2026

Muse Spark 1.2 placed on Artificial Analysis Pareto frontier at $0.40 per task, with Sol among compared models

6 August 2026

ARC-AGI-3 score jump from two API settings first disclosed

Sources

Rohan Paul Twitter