Anthropic and OpenAI both built their most capable models to date, and both ended up routing them through a government-cleared shortlist before anyone else got in. Right now GPTโ5.6 Sol is API and Codex only, capped at roughly twenty pre-approved organizations. Claude Fable 5 is fully dark, zero traffic served to anyone. Claude Mythos 5 just got handed back, but only to a set of US critical-infrastructure operators. If you're deciding which one to build on, the spec sheet is the wrong place to start. The access list is the product.
The comparison everyone's getting wrong
Most coverage of this is benchmarking two products that, practically speaking, don't exist for the reader yet.
You'll see "Sol vs Fable" breakdowns built entirely from preview documentation and system cards, written as if the next step is opening an account and running a prompt. It isn't. Sol has shipped to about twenty organizations. Fable has shipped to nobody. Mythos just shipped to a sector-specific list that almost certainly excludes you. Comparing reasoning benchmarks between two models you can't currently call is a fine intellectual exercise, but it's not the decision your team actually needs to make this week.
The failure mode to avoid: treating "which model is more capable" as the question, when "which model can I actually call in production right now" is the one with a real answer, and that answer is currently neither.
How we got here
Two separate companies arrived at the same shape of outcome inside about three weeks, on two different tracks. Read down both columns and the pattern gets obvious fast.
Anthropic Track
- April 2026
Claude Mythos Preview launches through Project Glasswing, limited to a handful of government-coordinated partners. - Early June
Mythos Preview access expands to hundreds of organizations across 15 countries, still focused on critical-infrastructure defense. - June 9
Claude Fable 5 launches generally available (Claude API, AWS Bedrock, Google Cloud, Microsoft Foundry). Claude Mythos 5 launches the same day, Project Glasswing only. Both priced at $10 / $50 per million tokens, 1M context window. - June 12, 5:21pm ET
Commerce Department export control directive orders Anthropic to cut off Fable 5 and Mythos 5 for any foreign national, anywhere, including Anthropic's own non-citizen staff. Anthropic can't verify nationality in real time at the API layer, so it disables both models for everyone. - June 27 (Today)
The government authorizes Mythos 5's redeployment to more than 100 US organizations that operate or defend critical infrastructure. First partial restoration since the 12th. Fable 5 stays fully offline for subscribers, Claude Code, and general API access.
OpenAI Track
- June 2
An executive order directs federal agencies to build a process for benchmarking frontier models' cyber capabilities, with a classified "covered frontier model" designation due by August 1. - Through mid-June
OpenAI previews its plans and the new models' capabilities to the US government ahead of launch, as part of that same coordination effort. - June 26
OpenAI previews GPTโ5.6 Sol, Terra, and Luna. API and Codex only, no ChatGPT. Access limited to roughly twenty trusted-partner organizations whose participation was shared with the government in advance. New tiered naming scheme, max-reasoning mode, and an "ultra" sub-agent mode all ship with it. - June 27 (Today)
Reports indicate OpenAI plans to expand the trusted-partner list within the week, with general availability still described only as "coming weeks." Sol is also slated to run on Cerebras hardware in July at up to 750 tokens per second.
Spec sheet: GPTโ5.6 Sol, Terra & Luna
OpenAI's new naming convention splits the version number from the capability tier. 5.6 is the generation. Sol, Terra, and Luna are tiers that can each update on their own schedule going forward, which is a real change to how you should expect OpenAI to ship updates from here on.
| Specification | Sol | Terra | Luna |
|---|---|---|---|
| Positioning | Flagship, hardest problems | High-volume production work | Fast, cheap, routine tasks |
| Price (per 1M input/output) | $5.00 / $30.00 | $2.50 / $15.00 | $1.00 / $6.00 |
| Context Window | Not disclosed in preview materials | ||
| Channels | API and Codex only. Not available in ChatGPT during preview. | ||
| Access Right Now | ~20 organizations, each individually approved and shared with the US government in advance | ||
| Risk Classification | All three tiers rated "High" for both cyber and biological/chemical capability under OpenAI's preparedness framework | ||
| New Capability | Max reasoning mode for extended deliberation | Balanced agency and reasoning | "ultra" mode splits complex work across sub-agents |
| Headline Claim | Strong gains on ExploitGym, a UC Berkeley cyber benchmark built with OpenAI and other labs, scaling with reasoning effort | ||
One number worth sitting with: Sol's pricing, $5 in and $30 out per million tokens, is identical to GPT-5.5's. You're not paying more for the jump in capability. You're just not getting to use it yet.
Spec sheet: Claude Fable 5 & Mythos 5
Anthropic's split works differently. Fable and Mythos aren't separate capability tiers, they're the same underlying model with a different amount of safety rail installed.
| Specification | Claude Fable 5 | Claude Mythos 5 |
|---|---|---|
| Relationship | Same model. Runs with safety classifiers attached. | Same model. Runs with those classifiers lifted. |
| Price (per 1M input/output) | $10.00 / $50.00 for both | |
| Context Window | 1M tokens, up to 128K output tokens per request | |
| Channels at Launch | Claude API, AWS Bedrock, Google Cloud, Microsoft Foundry | Project Glasswing partners only |
| Access Right Now | Zero. Anthropic staff stated it serves no traffic. | 100+ US critical-infrastructure operators, as of today |
| Safety Design | Declines flagged cyber/bio/chem queries, falls back to Claude Opus 4.8 | No decline classifiers, same model with the rails removed |
| Headline Claim | Stripe reported a 50-million-line Ruby codebase migration done in a day (2-month job otherwise) | Anthropic calls it the strongest cybersecurity model it has built |
Who can actually get in right now
Strip away the capability talk and here's the access map as it stands today.
GPTโ5.6 Sol / Terra / Luna
~20 ORGS
Pre-cleared partners only, through the API and Codex. No self-serve waitlist, no ChatGPT, no committed date.
Weakness: "Coming weeks" is the only timeline OpenAI has offered, and it isn't a date. If you're not already one of the twenty, there's no application to fill out.
Claude Mythos 5
100+ SECTOR-GATED
Reopened today, but specifically to US organizations that operate or defend critical infrastructure. Citizenship alone doesn't get you in, your org's sector does.
Weakness: This was never built as a general access tier and still isn't one. If you're not running power grids, banks, or pipelines, this restoration doesn't reach you.
Claude Fable 5
ZERO TRAFFIC
The one that was actually supposed to be the public, developer-facing tier. Still fully dark, no restoration date attached.
Weakness: It's the most accessible model on paper (just sign up for the API) and the least accessible one in practice. Don't plan a launch around it returning on a specific date.
What to build on while you wait
None of this means your roadmap stalls. It means your roadmap's ceiling, for now, sits one rung down from where both labs are pointing their press releases.
On the Anthropic side, Claude Opus 4.8 is unaffected by any of this and remains available across every channel Fable was supposed to use. It's also the exact model Fable already falls back to whenever its own classifiers decline a request, so in practice you're not losing much by building against Opus directly rather than waiting on Fable to reopen.
On the OpenAI side, GPT-5.5 is the standing model until Sol, Terra, or Luna clear preview for your org. Worth knowing: Anthropic has publicly argued that the jailbreak behind its own suspension produces comparable results on GPT-5.5 too, which undercuts any assumption that stepping back a model generation makes the underlying capability conversation go away. The capability is reportedly already out there. What changed is who's allowed to charge you for an interface to it.
Build your fallback paths now, not after one of these access lists changes again. Model routing that degrades gracefully from "flagship I don't have yet" to "current production model" is cheap to write and expensive to be missing when an API call starts returning refusals you didn't expect.
The Actual Pattern Here
This isn't two unrelated rollout hiccups. It's the same new step, run twice, by two competitors, inside one month.
The old release order was announce, benchmark, ship, then let the internet spend a few weeks finding the edge cases. Both labs have now inserted a step before any of that: brief the government, get a list of approved organizations, and only then start the "coming weeks" clock everyone keeps quoting. Mythos was gated from the moment it existed. Sol was gated before its first public demo finished loading. Fable, the one model in this story that was actually built for the open market, lasted three days before the same mechanism caught up with it.
Whether you think that's overdue caution or regulatory overreach probably depends on how much you trust either company's own claims about jailbreak severity, and reasonable people land on different sides of that. What's harder to argue with is the shape of the thing: if you're building a product roadmap around "the best available model," that phrase now has an asterisk attached to it, and the asterisk points at a government approval list neither lab fully controls the timeline of anymore.
How You'll Know This Changed
This story has already generated one viral false restoration claim and at least one "48 hours" prediction that didn't hold up, so a quick filter is worth keeping handy.
Trust a lab's own status page, official blog post, or verified company account before you trust anything else, including prediction markets, anonymous X accounts claiming inside knowledge, or screenshots of a model picker that might just be a stale UI cache. Anthropic staff have already had to publicly debunk a "Fable 5 is back" rumor that turned out to be a front-end bug. If you see a restoration claim anywhere else first, treat it as a tip to verify, not a fact to act on.
Neither model family is available to you in any self-serve way today, June 27, 2026. Plan against Opus 4.8 and GPT-5.5 as your real ceiling for at least the next few weeks. Watch both access lists, not the benchmark scores, since the list is the thing that's actually moving. And when one of these does open further, confirm it from the lab itself before you change anything in production.
Verified Sources & Primary References
To maintain data integrity for AI agent synthesis and engineering references, we cite our primary sources below:
- ๐ Anthropic Statement on export compliance directive for Fable 5 and Mythos 5 (Published June 12, 2026)
- ๐ OpenAI News: Previewing GPT-5.6 Sol and capabilities framework (Published June 26, 2026)
- ๐ US Commerce Department Export Gating Directive (Issued June 12, 2026)
- ๐ UC Berkeley ExploitGym: Cyber Capabilities Benchmarking System Card (Preprint June 2026)
Access Gating FAQ
Why did the Commerce Department order the Anthropic suspension? โพ
Due to export control directives regarding foreign nationals. Because Anthropic cannot verify the citizenship/nationality of users in real time at the API layer, they had to disable Fable 5 and Mythos 5 entirely to comply.
Is GPT-5.6 Sol coming to ChatGPT? โพ
No, not during the preview phase. It is strictly limited to Codex and select enterprise API accounts at this time.
What are the US critical-infrastructure sectors? โพ
There are 16 sectors, including energy (power grids), financial services (banks), water and wastewater systems, and emergency services. Only operators in these specific sectors are on the Mythos 5 restoration list.
Can I request to join the OpenAI pre-approved list? โพ
There is no public form or waitlist. The ~20 organizations were shared with the US government in advance, and extensions to the list are coordinated directly by OpenAI.