Written by AI agents, curated and verified by me.
GPT-5.6 generally available: the table says less than the text above it
- OpenAI
- Coding Agents
- Verification
OpenAI has made the GPT-5.6 family generally available: Sol as the flagship, Terra as a balanced model for everyday work, Luna as the cheapest tier. All three are available now in ChatGPT, in Codex, and through the API, with a global rollout that runs toward full availability over the following 24 hours. At the preview in late June, OpenAI explicitly said an expanded set of results would follow at general availability. It is here now. That is the interesting part, not the ranking: the coding claim can finally be read against the vendor’s own table.
What is new at general availability?
Mostly access and billing. In ChatGPT, Plus, Pro, Business, and Enterprise users reach Sol from the medium effort setting upward, with Pro and Enterprise also able to select GPT-5.6 Sol Pro. In ChatGPT Work and Codex, Free and Go users get Terra, while Plus and above can choose among Sol, Terra, and Luna with a per-model effort level. The max setting is available to everyone with GPT-5.6 access there and can be toggled on in settings. ultra is limited to Pro and Enterprise in ChatGPT Work, and to Plus and higher in Codex. Pricing per million tokens: Sol at 5 dollars input and 30 dollars output, Terra at 2.50 and 15 dollars, Luna at 1 and 6 dollars. Prompt caching becomes more predictable, with explicit cache breakpoints and a 30-minute minimum cache life. From this generation on, cache writes bill at 1.25 times the uncached input rate while cache reads keep the 90 percent discount. Two additions in the Responses API matter for agent work: Programmatic Tool Calling, where the model writes and runs small in-memory programs to coordinate tools and filter intermediate results, and Multi-Agent in beta, where one request runs concurrent subagents and synthesises their work.
What do the coding numbers actually say?
Sol leads, but not everywhere. On the Artificial Analysis Coding Agent Index v1.1, Sol scores 80 against 77.2 for Claude Fable 5, and OpenAI adds that it uses less than half the output tokens, takes less than half the time, and costs about a third less. The neighbouring rows are worth reading: Terra reaches 77.4, marginally above Fable 5, and Luna reaches 74.6, above Claude Opus 4.8 at 72.5. On Terminal-Bench 2.1, Sol stands at 88.8 and, with ultra, at 91.9, against 88.0 for Claude Mythos 5. On DeepSWE v1.1, Sol reaches 72.7 against 69.7 for Fable 5.
And then the same table lists SWE-Bench Pro. There Sol scores 64.6 percent, Claude Mythos 5 scores 80.3 percent, Fable 5 scores 80 percent, and Claude Opus 4.8 scores 69.2 percent. On that benchmark the flagship sits not only behind the competition but behind the competition’s previous generation. The Artificial Analysis Intelligence Index v4.1 runs the same way: Sol at 58.9, Fable 5 at 59.9. OpenAI publishes this itself and argues on time and cost, 61 percent less time at roughly half the estimated cost. That is a fair argument. It is simply a different argument from “best coding model”. Anyone making a purchasing decision should know which of the two is on the table.
Where the prose and the table diverge
In two places, and both are instructive. In the running text, OpenAI cites a new high of 53.6 on Agents’ Last Exam and a 13.1 point lead over Fable 5. The table lists 52.7 percent for Sol and 40.5 percent for Fable 5. The likely explanation is a different reasoning effort, except the table does not label one. Second, BrowseComp: the 92.2 percent quoted in the text appears in the table as the Sol Ultra column, meaning the four-agent configuration. Sol on its own sits at 90.4 percent. None of this is hidden, it is further down the same page. But it is the difference between the number you get quoted and the number you actually ordered.
What ultra means in production
max gives the model more time than xhigh to reason, explore alternatives, run checks, and revise its approach. ultra goes further and coordinates four agents in parallel by default, trading higher token use, according to OpenAI, for stronger results in less time. For BrowseComp and SEC-Bench Pro the company also shows 16-agent configurations. The consequence for your pipeline is sober: four parallel workstreams produce one result, and whoever signs it off sees that result, not the four paths to it. More parallelism increases the number of decisions nobody reads along with. The review effort does not disappear, it moves to the end.
More capability, considerably more blocking
In cybersecurity the jumps are large. On ExploitBench, OpenAI reports 73.5 percent against 47.9 percent for GPT-5.5 at a comparable output token budget; on SEC-Bench Pro, 71.2 against 45.8 percent; on ExploitGym, a rise in peak pass rate from 15.1 to 24.9 percent under a two-hour cap and to 33.7 percent with six hours. By OpenAI’s own testing, the models cross the Critical threshold in neither cyber nor biology. Before launch, roughly 700,000 A100-equivalent GPU hours went into automated black-box red teaming. The operationally important figure is a different one: Sol’s cyber safeguards block roughly ten times more potentially harmful activity than previous models. OpenAI names the resulting friction for legitimate work explicitly and offers an option in ChatGPT and Codex to retry a prompt on a lower-capability model. If you automate security work, plan for that case rather than discovering it during an incident.
Where GPT-5.6 earns its place
On long, tedious stretches with many tool calls, where Programmatic Tool Calling holds the intermediate results together instead of routing every tool response back through the model. The economically interesting part is less Sol’s peak numbers than Terra and Luna, which land close to Fable 5 on the coding index at half and a fifth of Sol’s output price. Test that against your own tasks, not against SWE-Bench Pro and not against an index. The table is public now, which is real progress over the preview, but it does not replace a test run in your repository. And the points where a human signs off stay the same: before the merge, before the deploy, before the migration. Reliability does not come from the model but from the architecture around it, as described in agentic engineering.