GPT-5.6 Sol: OpenAI’s New Flagship and the Price Cut Aimed at Claude

On 9 July 2026, after almost a fortnight of political choreography, OpenAI put the GPT-5.6 family on general release. The flagship, GPT-5.6 Sol, now powers the deeper reasoning settings in ChatGPT for paid subscribers, and it arrives with pricing designed to hurt: $5 per million input tokens and $30 per million output. That is exactly half what Anthropic charges for Claude Fable 5 on input, and 40% less on output, according to Anthropic’s own pricing page.

The launch came in two stages, and the first one was unusual. For two weeks the only people allowed near the models were partners the US government had approved by name. The same political pressure that took Claude Fable 5 offline in June shaped how GPT-5.6 reached the public.

Plenty of the early coverage is recycled press release. Here is what actually checks out against primary sources, and what it changes for anyone signing off a marketing budget this year.

What Is GPT-5.6 Sol?

GPT-5.6 Sol is the most capable model in OpenAI’s new GPT-5.6 series, sitting above two smaller siblings: Terra, a mid-priced model for everyday work, and Luna, the fastest and cheapest of the three. The release also introduces a new naming system. The number tells you the generation, while Sol, Terra and Luna are durable capability tiers that can be updated on their own schedules. Expect those three names to stick around.

The spec sheet reads like a machine built to be left alone with a job. The OpenRouter model listing puts the context window at 1,050,000 tokens, with output capped at 128,000. Call the plain gpt-5.6 alias in the API and Sol is what answers. Two new settings define how far you can push it. A max reasoning effort gives the model more thinking time than any previous setting, and an ultra mode coordinates four agents in parallel by default, splitting a hard task into concurrent workstreams and merging the results. Developers get a version of the same idea through a multi-agent beta in the Responses API, along with a feature called Programmatic Tool Calling that lets the model write small programs to manage its own tool use rather than routing every intermediate step back through the model.

Little of that sounds like marketing, granted. But the tools your team already pays for sit on these APIs, and OpenAI has plainly decided the metric that matters is finished work per pound spent. When that number moves, your software bills tend to move a few months later.

Why Was GPT-5.6 Held Back for Two Weeks?

The preview opened on 26 June 2026 to a handpicked set of partners, each of them named to the US government before getting access. TechCrunch reported that the restriction came at the administration’s request, following an executive order in June that asked AI companies to submit their most powerful models for government review before public release. During the preview, the models were only reachable through the API and Codex, and only for those approved partners.

OpenAI did not pretend to be happy about it. The company said publicly that government-gated access should not become the default for model releases, arguing it keeps useful tools away from the developers and defenders who need them. The caution was understandable. Weeks earlier, Anthropic had been forced to suspend Claude Fable 5 and Mythos 5 entirely under a government directive, a saga we covered in detail in our Claude Fable 5 analysis. OpenAI clearly wanted a staged release rather than a shutdown.

The gate lifted quickly. On 8 July, CNBC reported that OpenAI would release the models publicly that Thursday, and on 9 July general availability began rolling out worldwide across ChatGPT, Codex and the API over roughly 24 hours. Demand made itself felt almost immediately. Within days, OpenAI temporarily relaxed usage limits on Sol after a surge in the 48 hours following launch.

How Good Is GPT-5.6 Sol?

The honest answer is very good, with caveats. Every figure below comes from OpenAI’s own launch post, so treat them as a vendor’s best case until independent testing settles.

Coding is the headline. On the Artificial Analysis Coding Agent Index, Sol with max reasoning scores 80, which OpenAI says is a new state of the art and 2.8 points above Claude Fable 5, achieved with fewer than half the output tokens. On Terminal-Bench 2.1, which tests command-line work, Sol scores 88.8% and rises to 91.9% in ultra mode. For knowledge work, OpenAI reports a new high of 53.6 on Agents’ Last Exam, an evaluation of long professional workflows across 55 fields, beating Claude Fable 5 in its adaptive reasoning mode by 13.1 points. On the broader Artificial Analysis Intelligence Index, Sol lands within a single point of Fable 5 (58.9 against 59.9) while finishing tasks in 61% less time at roughly half the estimated cost.

The agentic browsing numbers deserve attention from anyone who cares about search. Sol scores 90.4% on BrowseComp, a benchmark of tasks that require researching the live web, and 92.2% in ultra mode. On OSWorld 2.0, which measures how well a model operates a real computer, it reaches 62.6%, passing Claude Opus 4.8 while using 85% fewer output tokens.

Anthropic has not been swept off the board, though, and OpenAI’s own comparison tables show it. On SWE-Bench Pro, a hard software engineering benchmark, Claude Fable 5 scores 80% against Sol’s 64.6%. Fable 5 also edges Sol on HealthBench Professional, 60.9% to 60.5%. The fair summary is that Sol competes at the frontier on most measures and wins decisively on cost per result, while Fable 5 keeps the crown on some of the most demanding engineering work.

There is one more area worth flagging. OpenAI describes GPT-5.6 as a step change in design judgement, able to inspect its own rendered output through computer use and fix visual problems before handing work back. Independent observers have started to agree: Tekedia reported that Sol has taken the top spot on the Design Arena benchmark. Early partners point the same way. Canva said the model was stronger than rivals for slide creation in its design evaluations and about 1.6 times more token-efficient, and Model ML found it produced more polished, presentable decks than Fable while using 39% fewer tokens per deck.

What Does GPT-5.6 Cost?

API pricing is straightforward across the three tiers, all in US dollars per million tokens:

For context, Claude Fable 5 sits at $10 input and $50 output on Anthropic’s price list, and OpenAI claims Terra and Luna can outperform Fable 5 on long professional workflows at around one sixteenth of the cost. Even allowing for vendor optimism, the gap in sticker price is real and it will pull the price of AI-powered tools down across the market.

Two billing details matter for anyone running the model at scale. Prompt caching now comes with explicit cache breakpoints and a 30-minute minimum cache life; cache writes cost 1.25 times the standard input rate while cache reads keep their 90% discount. And very long prompts carry a surcharge: OpenAI’s model documentation notes that requests above 272,000 input tokens are billed at twice the input rate and 1.5 times the output rate.

One caution on the headline numbers. Cheap tokens are not the same as cheap tasks. The max setting exists to burn more tokens on harder problems, and ultra runs four agents at once, which multiplies consumption by design. Price the job, not the token, and run a week of your own workloads through it before anyone signs anything.

Which ChatGPT Plans Include GPT-5.6 Sol?

According to OpenAI’s help centre, Plus subscribers get Sol in standard ChatGPT through the medium and high reasoning settings, while Pro, Business and Enterprise plans add the extra high setting and GPT-5.6 Sol Pro, a higher-capability option for difficult, longer-running tasks. GPT-5.5 Instant stays the default for quick everyday replies. Terra and Luna are not selectable in ordinary ChatGPT conversations.

The picture is different in ChatGPT Work and Codex. There, Free and Go users get Terra, while paid plans can choose between Sol, Terra and Luna and set an effort level for each. The max setting can be switched on by anyone with GPT-5.6 access in those products, and ultra is reserved for Pro and Enterprise in ChatGPT Work, or Plus and above in Codex.

For UK readers, the practical news is good. OpenAI confirms GPT-5.6 has no separate country list, so it is available wherever ChatGPT operates, including the UK. The rollout is gradual, which means the model may take a little while to appear in your account’s picker even on an eligible plan. And it will not stay confined to ChatGPT. Microsoft flipped Copilot’s preferred model to GPT-5.6 on launch day, so the model is already sitting inside Word, Excel and Outlook across a good chunk of British offices.

What Should Marketing Teams Actually Do about GPT-5.6?

Start with the money. Frontier-quality output at half the input price of its nearest rival changes the economics of every AI-assisted workflow an agency or in-house team runs, from content drafting and reporting pipelines to chatbots and internal research tools. If a supplier priced your AI tooling against 2025 token rates, this release is your renegotiation letter.

The design and document gains matter more than they first appear. Think about what actually eats a marketing week: deck production, template wrangling, the third round of visual tidy-up. A model that drafts an editable deck from a brief and then looks at its own rendered output to catch the ugly bits is pointed straight at that pile. It will not replace a designer with taste. It will compress the distance between a rough idea and a reviewable first draft, and the teams that fold that into their process will simply ship more.

The browsing benchmarks carry a quieter warning. A model that completes over 90% of live-web research tasks is a model that reads your website on behalf of your customers, and it will do so more often as these capabilities reach ChatGPT’s hundreds of millions of users. That raises the stakes for answer engine optimisation. Pages that state facts plainly, answer real questions under clear headings, and keep their information current get cited; pages that bury the answer do not. This is the same discipline we apply in our own AEO work, and GPT-5.6 just made it more valuable. 

Two cautions belong in any honest assessment. Every benchmark above is OpenAI’s own, published at launch, and independent results always land a little lower and a little messier. And the model’s cyber safeguards, which OpenAI says block roughly ten times more potentially harmful activity than previous models, can occasionally catch legitimate requests; there is a built-in option to retry on a lower-capability model when that happens. OpenAI’s system card also notes the model shows a somewhat greater tendency than GPT-5.5 to go beyond a user’s instructions in agentic coding tasks, at low absolute rates. Keep a human in the review chain. That rule outlives every model cycle.