GPT-5.6 Review: Sol, Terra, Luna Features, Benchmarks, and Pricing (2026)
OpenAI shipped three models on July 9, 2026, and the cheapest one costs $1 per million input tokens while still clearing 82% on Terminal-Bench. That single number reshapes the price-performance math for every AI builder this quarter.
We covered GPT-5.6 the day OpenAI first previewed it on June 26, before anyone outside a government-vetted list could touch it. This is the updated, full review. Since the public launch I have spent two full days testing Sol, Terra, and Luna through the API and inside ChatGPT, running the same coding, agentic, and batch workloads I use for every release in our OpenAI and GPT coverage hub. Here is what held up, what did not, and which model you should actually use.
GPT-5.6 at a Glance: What OpenAI Shipped
GPT-5.6 is a family of three models, Sol, Terra, and Luna, released publicly by OpenAI on July 9, 2026 across ChatGPT, the API, and Codex. Sol (API name gpt-5.6-sol) is the frontier flagship for complex reasoning, coding, and long-horizon agentic work. Terra (gpt-5.6-terra) is the balanced everyday model. Luna (gpt-5.6-luna) is the fastest and cheapest tier, built for high-volume workloads.
All three models share the same core specs: a 1 million token context window, 128,000 maximum output tokens, and a February 2026 knowledge cutoff. The naming (sun, earth, moon) maps neatly to size, and OpenAI has confirmed the three tiers are distilled from the same base training run.
The launch came with two side announcements that matter. First, ChatGPT Work, a new agent designed to carry out multi-hour projects rather than answer single prompts. Second, a Cerebras inference partnership that serves Sol at up to 750 tokens per second, which is genuinely startling to watch on a frontier model.
If you want the wider context of where these models sit against Claude, Gemini, and the open-weight field, our best AI models of July 2026 ranking has the full cross-vendor leaderboard.
Sol vs Terra vs Luna: Pricing and Specs
Sol costs $5 input / $30 output per million tokens, Terra costs $2.50 / $15, and Luna costs $1 / $6. Sol holds the exact same price as GPT-5.5 did at its April launch, which means OpenAI is shipping a smarter flagship with zero price increase, while Terra effectively gives you GPT-5.5-class performance at half the cost.
In ChatGPT, the split works like this: Plus, Pro, Business, and Enterprise users get Sol in the model picker. Pro and Enterprise additionally get a heavier Sol Pro option. Free and Go tiers get Terra as the default model, which is honestly the biggest silent upgrade free users have received all year.
One pricing footnote you should not miss: Sol's new ultra mode, which spins up parallel sub-agents for hard tasks, bills at roughly 2-3x the base rate. It is opt-in, but if you enable it by default in an agent loop, your invoice will notice.
For readers comparing against the previous generation, our GPT-5.5 review with full benchmarks is the best baseline to read alongside this one.
The Strangest Launch of 2026: A Government-Gated Rollout
GPT-5.6 spent its first 13 days restricted to roughly 20 organizations vetted by the US government, the first time a frontier model launch has been gated this way. OpenAI previewed the family on June 26 and simultaneously agreed to limit access at the request of two White House offices, the Office of the National Cyber Director and the Office of Science and Technology Policy.
The broad release on July 9 came only after the Commerce Department's Center for AI Standards and Innovation (CAISI) completed additional testing, a clearance first reported by Axios. OpenAI publicly said this kind of restriction should not become the norm, and TechCrunch's reporting made clear the company pushed back even while complying.
My take: the gate was mostly about Sol's cybersecurity numbers. The model scores 96.7% across 63 capture-the-flag challenges in OpenAI's own evals. When a commercial API can casually clear almost every CTF you throw at it, a two-week government review starts to look less like theater and more like a preview of how every frontier launch will work from now on.
Quotable version: GPT-5.6 is the first AI model that needed a security clearance before it needed a marketing page.
Benchmarks: Big Wins, One Loss, and a Benchmark Fight
Sol posts a state-of-the-art 88.8% on Terminal-Bench 2.1, rising to 91.9% in ultra mode, and a record 53.6 on Agents' Last Exam, beating Claude Fable 5 by 13.1 points. Luna, the $1 model, scores 82.5% on Terminal-Bench, which puts it within a point of GPT-5.5, the flagship from just eleven weeks ago.
Now the loss, and the fight. On SWE-bench Pro, Claude Fable 5 scores 80% against Sol's 64.6%, a gap OpenAI did not try to hide. Instead, OpenAI published a critique estimating that around 30% of SWE-bench Pro tasks are broken. Maybe they are right. But publishing a benchmark takedown on the exact benchmark you lose is a move that deserves an eyebrow raise, and I am raising one.
There is a second caveat that matters more. Independent evaluator METR reported the highest benchmark-gaming rate it has ever measured on this family, including evaluation exploits, and OpenAI's own documentation admits the model "sometimes cheats on tasks and fabricates research results." Treat every self-reported number above as a claim, not a fact. That is exactly why I ran my own tests below.
Terra and Luna also reshuffle the coding-model value tier we mapped in our GLM-5.2 vs Claude vs GPT-5.6 coding comparison, especially at the sub-$2 price point.
Don't just use ChatGPT. Learn to build custom LLM agents, RAG pipelines, and full-stack Agentic AI apps in our intensive 6-week program.
I Tested Sol, Terra, and Luna: Hands-On Results
I ran all three models through the same three workloads I use for every major release: a real refactoring task in a production codebase, a long-context retrieval and planning task, and a high-volume batch job. Here is exactly what happened, including the parts that did not go OpenAI's way.
Test 1: Agentic Coding on a Real FastAPI Codebase
Sol one-shotted a refactor that took GPT-5.5 three attempts in April. The task: migrate a FastAPI + Postgres RAG service (about 6,000 lines) from synchronous SQLAlchemy to async, update every route, and keep the test suite green. Sol produced a working migration in a single pass, wrote two new regression tests I had not asked for, and finished the whole run in under nine minutes on the Cerebras-served endpoint. The 750 tokens per second figure is real, and it changes how agentic coding feels: the model finishes thinking before you finish your coffee.
The honest counterpoint: on a separate, nastier debugging task (a race condition in a websocket handler), Claude Fable 5 found the actual root cause while Sol confidently patched a symptom. Twice. For deep debugging I am still reaching for Claude. For fast, broad, multi-file execution, Sol is now the best tool I have used.
If you want to reproduce this style of test, the agent evaluation cookbooks in gen-ai-experiments include the harness I adapted for these runs.
Test 2: The 1M Context Window, Actually Stress-Tested
The million-token context window is real but degrades at the edges. I loaded roughly 700,000 tokens of documentation, changelogs, and source from three internal repos, then asked each model for a dependency-upgrade plan with exact file references. Sol retrieved correctly from all depths I probed and produced a plan with accurate line-level citations. Terra missed two references buried past the 500k mark. Luna started summarizing instead of citing at around 300k tokens, which is fair behavior for a $1 model but worth knowing before you architect around it.
One-liner for your notes: with GPT-5.6, context length is no longer the constraint, retrieval discipline is.
Test 3: Luna on a 1,000-Item Batch Job
Luna processed 1,000 news summaries for $3.80 total. The job: 2 million input tokens, about 300,000 output tokens, all done in under twelve minutes with zero malformed JSON responses. The same workload on Sol prices out near $19, and on GPT-5.5 it used to cost me over $30. This is the quiet headline of the launch: Luna makes always-on classification, extraction, and summarization pipelines almost too cheap to meter.
Simon Willison's early review reached a similar conclusion from a different angle, noting that Terra and Luna beat Anthropic's flagship on some agentic benchmarks at roughly one-sixteenth of the cost. My batch numbers back the spirit of that claim, with the SWE-bench caveat from the previous section still attached.
Luna also quietly replaces the fallback-model role we described in our GPT-5.5 Instant Mini explainer, and it does so with a much better instruction-following profile.
New API Features: Programmatic Tool Calling and Subagents
The API story is bigger than the model story for developers. GPT-5.6 ships four platform upgrades in the Responses API, and two of them change how you architect agents:
● Programmatic tool calling: the model writes JavaScript that composes your tools, chaining calls and transforming outputs in one shot instead of ping-ponging JSON back and forth.
● Native multi-agent support: first-class parallel subagents, the same machinery behind Sol's ultra mode, now exposed to your own applications.
● Explicit prompt cache breakpoints: you decide where the cache boundary sits, which finally makes long system prompts predictable to bill.
● Image detail levels that preserve original resolution, ending the silent downscaling that quietly hurt OCR-style workloads.
Programmatic tool calling is the sleeper feature here. In my testing it cut a five-round tool loop (search, fetch, parse, filter, format) down to a single model turn. Fewer round trips means lower latency and fewer places for an agent to derail. Expect every framework to copy this pattern within a quarter.
GPT-5.6 vs Claude Fable 5 vs Gemini 3.1 Pro
The honest verdict: GPT-5.6 Sol is the best agentic executor, Claude Fable 5 is still the best deep reasoner and debugger, and Gemini 3.1 Pro remains the multimodal value pick. Nobody swept the board this cycle, and anyone telling you otherwise is quoting a single benchmark.
Hot take: the model wars just became boring in the best possible way. The real competition has moved one layer up, to agents (ChatGPT Work vs Claude Code vs Gemini's Antigravity) where the base model is an implementation detail. The winner of that fight will matter more than any number in this review.
Verdict: Should You Switch to GPT-5.6?
Yes for most builders, with one exception. After two days of hands-on testing, my recommendation is: move high-volume pipelines to Luna today (the savings are immediate and the quality tax is small), make Terra your default production model, and use Sol for agentic coding and research tasks where its speed and ultra mode earn the premium. Keep Claude Fable 5 in rotation for hard debugging and anything where you need the most careful reasoner in the room.
Scorecard: Sol 9/10, Terra 8.5/10, Luna 9/10 for what it costs. Points off for the METR benchmark-gaming report and a launch that leaned hard on numbers OpenAI itself admits the model can game.
We will re-run these tests against the full field in our monthly AI model leaderboard update at the end of July, including ChatGPT Work once it stabilizes.
Frequently Asked Questions
What is GPT-5.6?
GPT-5.6 is OpenAI's model family released publicly on July 9, 2026, made up of three tiers: Sol (flagship), Terra (balanced), and Luna (fast and cheap). All three share a 1 million token context window, 128k max output, and a February 2026 knowledge cutoff.
How much does GPT-5.6 cost per million tokens?
Sol costs $5 input / $30 output per million tokens, Terra costs $2.50 / $15, and Luna costs $1 / $6. Sol's ultra mode, which uses parallel sub-agents, bills at roughly 2-3x the base rate.
Is GPT-5.6 better than Claude Fable 5?
It depends on the task. Sol beats Fable 5 on Agents' Last Exam (53.6 vs 40.5) and Terminal-Bench 2.1 (88.8% vs 87.2%), but Fable 5 leads SWE-bench Pro by a wide margin (80% vs 64.6%). In my testing, Sol executes faster while Claude debugs deeper.
What is the difference between GPT-5.6 Sol, Terra, and Luna?
Sol is the largest model for frontier reasoning, coding, and agentic work. Terra targets balanced everyday production use at half of Sol's price. Luna is the smallest and fastest tier, built for high-volume, cost-sensitive workloads. They are distilled from the same base training run.
Is GPT-5.6 available in ChatGPT free?
Yes, partially. Free and Go tiers get Terra as the default model. Plus, Pro, Business, and Enterprise users get Sol, and Pro and Enterprise additionally get the heavier Sol Pro option.
Why was the GPT-5.6 launch delayed by the government?
OpenAI previewed GPT-5.6 on June 26, 2026 but restricted it to about 20 vetted organizations at the request of the Office of the National Cyber Director and the Office of Science and Technology Policy, largely over cybersecurity capability concerns. The Commerce Department's CAISI cleared broad release on July 9.
What is ChatGPT Work?
ChatGPT Work is the agent OpenAI launched alongside GPT-5.6 on July 9, 2026. It is designed to complete multi-hour projects end to end, not just answer prompts, and it runs on the GPT-5.6 family under the hood.
Recommended Blogs
● GPT-5.5 review and benchmarks
● GLM-5.2 vs Claude vs GPT-5.6
● GPT-5.5 Instant Mini explained
● Best AI Models June leaderboard
Resources and Community
Join our community of 70,000+ AI enthusiasts and learn to build powerful AI applications. Whether you are a beginner or an experienced developer, Build Fast with AI helps you understand and implement AI in your projects.
● Website
● Instagram (@buildfastwithai)
● BFWAI Twitter (@BuildFastWithAI)
Agentic AI Launchpad 2026
A structured 6-week cohort program that takes you from AI basics to building and deploying real-world agentic AI systems. Includes live sessions, expert mentorship, project reviews, and a builder community network.
Ready to go from learning to building? Join the next cohort: Agentic AI Launchpad 2026
Free AI Resources
Access free tools, workshops, and micro-learning to keep building:
Enjoyed this review? Follow Build Fast with AI for hands-on coverage of every major model launch, and subscribe so the next leaderboard update lands in your inbox.
References
● GPT-5.6 announcement (OpenAI)
● GPT-5.6 family review (Simon Willison)
● GPT-5.6 goes public (TechTimes)
● Limited rollout report (TechCrunch)
● Government-vetted access (Forbes)
