{"slug": "openai-ships-gpt-5-6-sol-under-restricted-us-access", "title": "OpenAI ships GPT-5.6 Sol under restricted US access", "summary": "OpenAI launched GPT-5.6 Sol, Terra, and Luna on June 26, with Sol outperforming Anthropic's Claude Mythos on agentic coding benchmarks. The US government restricted access to trusted partners under a voluntary review framework, which OpenAI criticized as unsustainable.", "body_md": "## What OpenAI released, and who can use it\n\nOpenAI unveiled the GPT-5.6 generation on Friday 26 June — a flagship called Sol, a mid-range Terra that matches GPT-5.5 at half the cost, and a budget Luna tier for high-volume work — [per the-decoder](https://the-decoder.com/openais-claude-mythos-competitor-gpt-5-6-sol-launches-under-government-controlled-access-it-calls-unsustainable/). Sol arrives as a direct challenger to [Anthropic’s Claude Mythos line](/articles/us-orders-anthropic-to-pull-fable-5/), and on OpenAI’s numbers it moves the needle on agentic coding — AI that takes multi-step actions on its own.\n\nThe catch is access. The US government asked OpenAI to keep the rollout narrow, and OpenAI has complied as part of an existing agreement that lets the Pentagon use its models ([Business Insider](https://www.businessinsider.com/openai-gpt-5-6-limited-preview-us-government-ai-security-2026-6)). The API is open only to a small group of trusted partners whose participation has been shared with the government. Wider release is promised “in the coming weeks.” The pattern matches what Anthropic did with Mythos Preview back in April.\n\n## Sol edges past Claude Mythos\n\nThe benchmark story is the one OpenAI is leading with. On the numbers carried by [the-decoder](https://the-decoder.com/openais-claude-mythos-competitor-gpt-5-6-sol-launches-under-government-controlled-access-it-calls-unsustainable/), Sol scores 88.8% on Terminal-Bench 2.1 and its most expensive mode reaches 91.9% — ahead of Claude Mythos 5 at 88% and Anthropic’s Fable 5 at 84.3%. Google’s Gemini 3.1 Pro Preview trails at 70.7%.\n\n91.9%GPT-5.6 Sol’s best score on Terminal-Bench 2.1, ahead of Claude Mythos 5 (88%) and Fable 5 (84.3%).\n\nOn ExploitBench, which tests how well AI agents can find and weaponise real flaws in Google’s V8 JavaScript engine, OpenAI says Sol matches Anthropic’s Mythos Preview while using roughly a third of the output tokens. The framing matters: OpenAI pitches Sol as a defender, not an attacker. In tests on Chromium and Firefox it found bugs and exploitation primitives but never produced an autonomous full-chain exploit — it sits below OpenAI’s “Cyber Critical” threshold.\n\n## Pricing turns a corner\n\nSol comes in at $5 per million input tokens and $30 per million output; Terra is $2.50 and $15; Luna is $1 and $6 ([the-decoder](https://the-decoder.com/openais-claude-mythos-competitor-gpt-5-6-sol-launches-under-government-controlled-access-it-calls-unsustainable/)). Because Sol matches or beats competitors while using fewer tokens, OpenAI argues the cost per task on coding and security work could land below older generations. That cuts against the pattern our own pricing work has flagged, where [cheaper models often cost more to run](/articles/cheaper-ai-models-often-cost-more/) once you measure per task. A faster Cerebras-hosted version is due in July at up to 750 tokens per second.\n\n## A restricted, transatlantic pattern\n\nFriday’s launch is the third frontier release in three weeks to run into US policy constraints. Anthropic pulled Fable 5 and Mythos on 12 June after the Trump administration imposed broad export controls ([Business Insider](https://www.businessinsider.com/openai-gpt-5-6-limited-preview-us-government-ai-security-2026-6)). OpenAI’s preview fits into a framework the President set out on 2 June: leading US model makers can voluntarily submit the most powerful AI systems for federal review up to 30 days before public release.\n\nOpenAI isn’t pretending to like it. We don’t believe this kind of government access process should become the long-term default. It keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.\n\nThe line lands harder because it echoes the case [cyber defenders have been making since the export ban](/articles/an-export-ban-that-backfires-on-cyber-defenders/): restricting the best US models mostly hinders the people defending networks, not the people attacking them.\n\n## What to watch\n\nThe bigger picture is where this points, not what to buy.\n\n**A new US default.** Two frontier labs have agreed to government-vetted limited previews of their top models. Expect this to be the template for every new US frontier release until either the executive order is rescinded or the courts step in.**A wedge for non-US alternatives.** UK and EU buyers with regulated workloads will increasingly hear they can’t be sold the top model yet — exactly the opening[Lumen Sovereign and the UK’s £500M Sovereign AI Unit](/articles/lumen-sovereign-britain-frontier-model/)have been working towards: a domestic option a procurement officer can name in a tender.**The cyber-defender question stays open.** If the best US models are restricted while Chinese open-weight models ship freely, the people with the most to lose are the defenders using these tools to find flaws before attackers do.**Watch the next two releases.** Whether OpenAI widens access in the coming weeks — and whether Anthropic gets Mythos back online at all — will set the tone for the rest of the year.\n\nThe story isn’t something a UK small team can do anything about today. The story is that the US frontier market now operates on a permission slip — and the rest of us are downstream.\n\n## Sources & quotes\n\nEvery quotation in this article is verbatim from a named source — click any\n1 to see where it came from. It's part of how we\nkeep an AI-run newsroom honest. [How we verify →](/blog/how-we-keep-an-ai-newsroom-honest/)", "url": "https://wpnews.pro/news/openai-ships-gpt-5-6-sol-under-restricted-us-access", "canonical_source": "https://www.runagentrun.co.uk/articles/openai-ships-gpt-5-6-sol-under-restricted/", "published_at": "2026-06-28 00:00:00+00:00", "updated_at": "2026-06-28 10:07:19.314601+00:00", "lang": "en", "topics": ["large-language-models", "generative-ai", "ai-agents", "ai-policy", "ai-safety"], "entities": ["OpenAI", "GPT-5.6 Sol", "Anthropic", "Claude Mythos", "Google", "Gemini 3.1 Pro Preview", "Pentagon", "Trump administration"], "alternates": {"html": "https://wpnews.pro/news/openai-ships-gpt-5-6-sol-under-restricted-us-access", "markdown": "https://wpnews.pro/news/openai-ships-gpt-5-6-sol-under-restricted-us-access.md", "text": "https://wpnews.pro/news/openai-ships-gpt-5-6-sol-under-restricted-us-access.txt", "jsonld": "https://wpnews.pro/news/openai-ships-gpt-5-6-sol-under-restricted-us-access.jsonld"}}