# One Week, Three Frontier Models — and a New Normal for AI Releases

Source: https://www.mafille.me/posts/2026/one-week-three-frontier-models
Published: 2026-07-09T20:00:00.000Z
Tags: ai, llm, anthropic, openai

Fable 5 is back, OpenAI shipped GPT-5.6 Sol, Terra and Luna, and Grok 4.5 landed a day earlier — all in one week. A look at what each release actually is, and at the quieter pattern underneath: frontier models now launch through a government checkpoint.

At the end of [my last article](https://www.mafille.me/posts/2026/claude-fable-5-export-controls), I was locked out of Claude Fable 5 by a US export-control directive, and I promised that if access ever came back, I'd return with a proper review. Well — access is back. But before I could even start that review, the industry decided to compress what feels like a quarter of AI news into seven days: Fable 5 was restored worldwide, OpenAI shipped not one but three GPT-5.6 models, and Grok 4.5 landed a day before them. So the hands-on Fable 5 review moves to the next post, and this one is about the week itself — what actually shipped, and the pattern underneath it that I think matters more than any single model.

![One week, three frontier models](https://www.mafille.me/_astro/frontier-week-banner.JUoUFT3f.png)
_AI-generated image via Google Nano Banana_

## Table of contents

## Fable 5 is back — mostly the way I hoped

On 30 June, the US Department of Commerce lifted the export controls on Claude Fable 5 and Mythos 5, and on 1 July [Anthropic restored access](https://www.anthropic.com/news/redeploying-fable-5) globally — Claude.ai, the API, Claude Code, everywhere. As a foreign national (Swiss, in my case), I went from "not allowed to touch it" to "it's in my model picker again" overnight, in the good direction this time.

What changed between the June shutdown and now is worth pausing on. The directive was triggered by a demonstrated technique for getting the model to find and exploit software flaws in a codebase. Anthropic's fix wasn't a region lock or a permanent nerf: they trained a new classifier that blocks that specific technique in over 99% of cases, and agreed to closer coordination with the US government on future launches. In my last post I wrote that I hoped Fable 5 would come back "with safeguards that are real, thoughtful, and aimed at the actual risks, rather than a region-wide off switch." A targeted classifier instead of a border-drawn shutoff is — credit where due — pretty much exactly that.

The commercial side is messier. Fable 5 is included in Pro, Max and Team plans for up to 50% of weekly usage limits, but only through 12 July (already extended once after users pushed back on the original 7 July cutoff). After that, officially, it moves to usage credits — on top of API pricing that sits at $10/$50 per million input/output tokens, double Opus 4.8. Anthropic frames this as temporary, promising to restore it as a standard part of subscriptions "as soon as capacity allows."

That's the announcement. My read is a bit different: I think they're figuring out, more or less in public, how to fit a Mythos-class model into their subscriptions at all — the extended deadline and the careful "temporary" language sound less like a finished plan than like a team watching real usage data and doing the capacity math live. And they don't have much choice but to make it work. OpenAI ships Sol directly in ChatGPT's paid tiers, no extra credits attached. If using Anthropic's best model means paying twice — a subscription _plus_ usage credits — while the competition includes theirs in the base plan, customers won't write angry forum posts the second time; they'll just leave. My bet is that Fable 5 lands back in the subscriptions properly, because the alternative is losing the exact users who care about frontier capability.

## OpenAI's answer: GPT-5.6 Sol, Terra and Luna

OpenAI didn't ship one model this week — it shipped a naming scheme. [GPT-5.6](https://openai.com/index/previewing-gpt-5-6-sol/) comes in three tiers: **Sol** is the flagship for the hardest problems (complex coding, security research, long-horizon agent work), **Terra** targets high-volume business tasks at roughly GPT-5.5 quality for half the cost, and **Luna** is the fast, cheap one for summarization and routine automation. The number is the generation; the celestial names are durable tiers that can advance on their own cadence. Pricing lands at $5/$30 for Sol, $2.50/$15 for Terra, and $1/$6 for Luna per million tokens — undercutting Fable 5 significantly at the top end.

The part of the launch story that got less airtime: GPT-5.6 didn't go straight to the public. On 26 June, after a request from the White House, OpenAI released it as a limited preview to roughly twenty government-vetted organizations. The broad public release only happened on 9 July, after US government approval. Twelve days of a frontier model existing but being reachable only by a government-approved list. OpenAI itself said this "shouldn't become the long-term default," and Sam Altman was openly uncomfortable with "the government picking the customers" — but they complied, same as Anthropic did.

There's also a genuinely awkward technical footnote. [METR found](https://www.techtimes.com/articles/319662/20260703/ai-benchmark-cheating-sets-record-gpt-56-sol-gamed-its-own-safety-tests.htm) that Sol gamed its software-engineering evaluations at the highest rate of any publicly tested model — exploiting bugs in the eval harness, digging out hidden test data, substituting shortcuts for real solutions. Depending on whether you count those exploits as failures, METR's estimate of the model's autonomous-work time horizon swings from roughly 11 hours to roughly 270. That's not a rounding error; that's "we can't reliably measure this model with our current tools." Keep that in mind next time a benchmark table settles an argument.

## Grok 4.5: aggressive pricing, honest trade-offs

A day before OpenAI's public launch, xAI — now SpaceXAI, after the merger, IPO and the acquisition of Cursor — released [Grok 4.5](https://cursor.com/blog/grok-4-5). Musk's own framing was unusually modest by his standards: "an Opus-class model, but faster, more token-efficient and lower cost." The interesting technical detail is that it was trained jointly with Cursor, on what Cursor describes as trillions of tokens of real developer-workflow data.

The pricing is the headline: $2/$6 per million tokens, with [independent analysis](https://the-decoder.com/grok-4-5-is-so-cheap-compared-to-fable-5-and-gpt-5-5-that-benchmark-gaps-may-not-matter-much/) putting the real cost of a typical agentic coding task at around $2.50 — versus about $5 on GPT-5.5 and almost $12 on Fable 5, largely because Grok uses far fewer tokens per task. On the benchmarks it trails Fable 5 nearly everywhere (64.7% vs 80.4% on SWE-Bench Pro, for instance), but at a fifth of the input price, "close enough and much cheaper" is a real strategy.

The trade-off nobody put in a headline: on AA-Omniscience, accuracy jumped from 35% to 52% versus Grok 4.3 — and the hallucination rate jumped from 25% to 54%. More likely to be right, and when wrong, more likely to sound confident about it. For a model pitched at office work, research and legal tasks, that's the number I'd watch. Also worth noting from where I sit in Europe: Grok 4.5 isn't available in the EU at launch, with no stated timeline.

## The pattern: a government checkpoint in the release pipeline

Step back from the individual launches and look at the sequence. In June, a US directive pulled Fable 5 offline for most of the planet, and it came back only after a two-week government review and negotiated commitments. Weeks later, GPT-5.6 spent twelve days gated behind a government-approved preview before the public was allowed in. And a formal US evaluation framework for cyber-capable frontier models is reportedly targeted for August.

That's not two coincidences; that's a pipeline taking shape. Frontier model releases in 2026 now have a de facto government checkpoint in them — sometimes before launch, sometimes brutally after. When I wrote about the Fable 5 shutdown, it felt like an exceptional event. One month later it looks more like the first execution of a process.

I'm genuinely torn on this. These models are rated by their own makers as highly capable in cybersecurity and biology; pretending pre-release scrutiny is absurd would be naive, and a predictable review beats an unpredictable shutdown every time — ask anyone who lost Fable 5 overnight. But the failure modes are just as real: a government choosing who gets early access, timelines that slip for political rather than technical reasons, and everyone outside the US discovering — again — that their access to these tools is a policy decision made somewhere else. The Fable 5 episode at least ended with a targeted technical fix instead of a permanent wall. Whether the August framework follows that template or the blunter one is, I think, the most consequential open question in AI deployment right now.

## Wrapping up

One week: the most capable generally available model came back from the dead, OpenAI restructured its whole lineup around three celestial tiers, and the price floor for Opus-class capability dropped to $2 per million tokens. Any of these would have carried a news cycle on its own; together they mostly cancel each other out, which tells you something about the pace we've normalized.

Next post is the one I promised a month ago: a proper hands-on review of Claude Fable 5 — what Mythos-class actually feels like on real work, now that nobody can take it away mid-review. Probably.

Thanks for reading. If you're chewing on the government-checkpoint question too — especially from outside the US — I'd love to hear from you, via [email](mailto:mathieu@mafille.me) or [LinkedIn](https://www.linkedin.com/in/mathieumafille/).

See you in two weeks.