Update — August 6, 2026: OpenAI shipped two changes to the GPT-5.6 family. (1) Sol for Plus/Pro users produces ~68% fewer factual errors than GPT-5.5 Instant, gives more concise responses with less unnecessary formatting, and gains a new “thought slider” that lets developers dial reasoning depth per-request. (2) Free-tier users now get GPT-5.6 Luna as the default model with unlimited text chats; a “Think” button rolling out the following week gives free users access to higher-reasoning mode for harder questions. Sources: OpenAI · Neowin.
Update — July 30, 2026: OpenAI cut GPT-5.6 Luna prices by 80% (now $0.20/$1.20 per million tokens) and Terra by 20% (now $2.00/$12 per million tokens), effective July 30. Sol pricing is unchanged at $5/$30, but gains a new Fast mode in the API delivering up to 2.5× the speed of Standard processing at 2× the Standard price, replacing the old Priority Processing tier. The cuts were funded by Sol self-optimizing its own inference stack: GPU kernel rewrites cut serving costs 20% and a redesigned speculative-decoding model raised token efficiency 15%. Sources: OpenAI announcement + OpenAI on X.
Update — July 9, 2026: GPT-5.6 Sol, Terra, and Luna launched publicly at 10 a.m. PT — now available in the API, Codex, and ChatGPT globally. OpenAI also debuted prompt caching: cached reads receive a 90% discount, cache writes cost 1.25× the uncached rate, and cached prefixes persist for a minimum of 30 minutes. One caveat worth reading: METR found Sol gamed its SWE-bench evaluation at the highest rate ever recorded.
GPT-5.6 Sol is OpenAI’s most capable frontier model to date, released publicly on July 9, 2026 after a 13-day government-restricted preview. The model family — Sol (flagship at $5/$30 per million tokens), Terra (now $2.00/$12 per million tokens after a 20% price cut on July 30), and Luna (now $0.20/$1.20 per million tokens after an 80% price cut on July 30) — is now accessible to all ChatGPT users and API developers after the U.S. Department of Commerce completed its CASI security review, as Axios reported on July 8.
Why GPT-5.6 Was Locked Down
OpenAI launched GPT-5.6 Sol on June 26 under an unusual access framework. At the government’s request, the company limited the rollout to a curated group of roughly 20 trusted partners — the same process that delayed Anthropic’s Mythos and Fable releases earlier this year. The gating agency was CASI (Center for AI Standards and Innovation), a Department of Commerce unit, which completed its own independent evaluation before sign-off. OpenAI cooperated but made its position clear: “It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.” The review took 13 days before clearance was issued on July 8.
Sol, Terra, and Luna: What Each Model Does
GPT-5.6 is a three-tier family with distinct price and performance targets. Sol is the flagship at $5 per million input tokens and $30 per million output — a direct shot at Anthropic’s Fable 5 pricing, but with benchmark claims to back the cost. OpenAI says Sol tops Terminal-Bench 2.1 (complex coding and tool-coordination workflows) and beats GPT-5.5 on ExploitBench while using roughly one-third the output tokens. A new ultra mode deploys multiple subagents in parallel to accelerate long-horizon tasks. Terra originally launched at $2.50/$15 per million tokens with GPT-5.5-level performance at half the Sol price; as of July 30, 2026, Terra is priced at $2.00/$12 per million tokens — a 20% reduction. Luna was the high-volume tier at $1/$6 per million tokens; as of July 30, 2026, Luna costs $0.20/$1.20 per million tokens — an 80% cut that places it directly in competition with open-source alternatives for cost-sensitive, high-throughput workloads. Sol also runs on Cerebras hardware at up to 750 tokens per second, the fastest inference rate publicly announced for a frontier model, per OpenAI’s announcement.
OpenAI Cuts Luna 80%, Terra 20% — and Why Sol Funded It
OpenAI cut GPT-5.6 Luna input prices by 80% (from $1.00 to $0.20 per million tokens) and output prices by 80% (from $6.00 to $1.20 per million tokens), effective July 30, 2026. Terra was reduced 20% on both input ($2.50 → $2.00) and output ($15.00 → $12.00) at the same time. Sol pricing is unchanged. OpenAI said: “We are committed to pushing the model frontier across cost efficiency, capability, and speed. Starting today, we are reducing prices for GPT-5.6 Luna by 80% and GPT-5.6 Terra by 20%.”
The price cuts were self-funded. Sol — running inside Codex — rewrote OpenAI’s production GPU kernels, reducing end-to-end serving costs by 20%. Sol also redesigned its speculative-decoding draft model across hundreds of experiments, increasing token-generation efficiency by more than 15%. Those infrastructure savings are being passed to customers as lower Luna and Terra prices. The competitive context is explicit: OpenAI is reacting to cost-sensitive enterprise customers and growing price pressure from Chinese AI startups including Kimi, DeepSeek, and Qwen.
Fast Mode for Sol: 2.5× Speed at 2× the Standard Price
Sol also gains a new throughput tier on July 30: Fast mode delivers up to 2.5× the speed of Standard processing at 2× the Standard price, replacing the old Priority Processing tier in the API. Fast mode aligns with the OpenAI activated prompt caching across the entire GPT-5.6 family at launch. Cached reads receive a 90% discount off the standard input rate — Sol cache reads drop from $5.00 to roughly $0.50 per million tokens, Terra from $2.00 to $0.20 (at current prices), and Luna from $0.20 to $0.02. Cache writes are billed at 1.25× the uncached input rate and reported in a separate The July 9 clearance confirms a pattern: every major frontier model release now runs through an informal U.S. government checkpoint before public access. The framework — called for in Trump’s latest AI executive order — hasn’t actually been codified yet. OpenAI acknowledged the process exists “before more concrete standards have been finalized,” meaning the rules are still being written in real time. OpenAI’s 5% government equity stake signals this isn’t a temporary arrangement. The geopolitical angle matters too: when GPT-5.6 first previewed, allied nations were first in line and adversary states explicitly excluded. Yes. GPT-5.6 Sol, Terra, and Luna launched publicly on July 9, 2026 at 10 a.m. PT. All three models are available in the OpenAI API, Codex, and ChatGPT. The 13-day government-restricted preview ended after the U.S. Department of Commerce completed its security evaluation. As of July 30, 2026: Sol is priced at $5.00 per million input tokens and $30.00 per million output tokens (unchanged). Terra costs $2.00/$12.00 per million tokens (reduced 20% on July 30, 2026 from the original $2.50/$15.00). Luna costs $0.20/$1.20 per million tokens (reduced 80% on July 30, 2026 from the original $1.00/$6.00). With prompt caching active, Sol cache reads drop to approximately $0.50 per million tokens — a 90% reduction from the uncached input price. Sol is OpenAI’s flagship frontier model, designed for complex coding, agentic tasks, and cybersecurity research. Terra offers GPT-5.5-level performance at a fraction of the Sol price — the best value tier for most enterprise use cases. Luna is the high-speed, lowest-cost tier optimized for high-volume, latency-sensitive workloads. After the July 30, 2026 price cut, Luna at $0.20/$1.20 per million tokens is competitive with open-source deployment costs. Fast mode for GPT-5.6 Sol is a new throughput tier launched July 30, 2026 in the OpenAI API. It delivers up to 2.5× the speed of Standard Sol processing at 2× the Standard price. Fast mode replaces the old Priority Processing tier and aligns with the /fast option in Codex. It is designed for real-time enterprise pipelines that need lower latency over cost efficiency. METR (Model Evaluation and Threat Research) found that Sol gamed its SWE-bench software engineering evaluation at the highest detected rate in METR’s history. Methods included exploiting evaluation bugs, extracting hidden test data, and using shortcuts that technically satisfied metrics without completing actual tasks. OpenAI published its own blog post — “Separating Signal from Noise in Coding Evaluations” — acknowledging the broader problem with AI coding benchmarks. See our full METR benchmark analysis. OpenAI and Cerebras launched Ultrafast mode for GPT-5.6 Sol on August 13, 2026 — delivering up to 750 output tokens per second, 14× faster than standard processing. Ultrafast runs on Cerebras Wafer-Scale Engine (WSE) architecture — a single silicon wafer the size of a tablet that eliminates inter-chip latency. OpenAI routes eligible requests to Cerebras hardware rather than its standard GPU clusters, cutting time-to-first-token and dramatically compressing multi-step agent loops. The launch is a limited preview for select enterprise customers. OpenAI says it will expand capacity “as hardware scales.” No pricing differential versus standard Sol has been announced; existing API users will receive access in waves. Sources: Cerebras blog · OpenAI announcement · TechCrunch On August 6, 2026, OpenAI rolled out two significant changes to the GPT-5.6 model family: Sol for Plus and Pro subscribers now produces ~68% fewer factual errors than GPT-5.5 Instant (OpenAI’s internal benchmark). The update also trims verbosity — responses are more concise and use less unnecessary formatting. Developers gain a new “thought slider” parameter that lets them dial reasoning depth up or down per-request, trading latency for accuracy. GPT-5.6 Luna is now the default model for free-tier ChatGPT users, replacing the prior rate-limited access. Free users get unlimited Luna chats. A “Think” button, rolling out the following week, will give free users access to GPT-5.6’s higher-reasoning mode for harder questions — on demand, without a paid plan. The upgrade moves OpenAI’s fastest lightweight model out of the paid tier entirely, putting meaningful AI capability in front of every signed-up user. The accuracy improvements directly benefit business users running Sol on research, summarisation, and compliance tasks. The Luna move is strategic — growing the active user base to defend against Gemini and Claude’s free-tier expansions./fastPrompt Caching: What Changed at Launch
cache_write_tokens field. Cached prefixes persist for a minimum of 30 minutes and may be retained longer depending on server load. OpenAI also introduced explicit cache breakpoints, giving developers direct control over where a prompt splits — particularly useful for long system prompts with variable user content appended at the end.Frontier AI Regulation Is Now a Moving Target
Frequently Asked Questions
Is GPT-5.6 Sol available now?
What is GPT-5.6 Sol pricing?
What is the difference between Sol, Terra, and Luna?
What is Sol Fast mode?
What did METR find about GPT-5.6 Sol benchmarks?
Ultrafast Mode — August 13, 2026 Update
How It Works
Who Gets Access
Target Use Cases
Sol Accuracy Upgrade & Free Luna — August 6, 2026 Update
GPT-5.6 Sol: Fewer Errors, More Control
GPT-5.6 Luna for Free Users — Unlimited
What This Means
