GPT-5.6, Gemini 3.5 Pro, Claude Mythos: What Shipped, What Slipped, and What Comes Next

Back in late June, nearly everything in this story was a rumor: leaked context windows, prediction-market odds, a government hold with no end date. Seven weeks later most of it has resolved, which makes this a rare chance to score the rumor mill against reality. GPT-5.6 shipped. Claude Fable 5 was pulled, then restored, then moved up-market. Gemini 3.5 Pro, promised for June, is still missing in mid-August. Here is the record, what each rumor turned out to be worth, and what the next wave (OpenAI's Astra, Google's Gemini 4) actually looks like. Then the practical part: should you wait for the next model before paying for anything?

The scorecard, in one paragraph

If you only read one paragraph: as of mid-August 2026, GPT-5.6 is publicly available. OpenAI launched it on July 9, 2026 in three tiers, Sol, Terra, and Luna, after the US Department of Commerce cleared the release. Claude Fable 5, suspended under an export-control directive three days after its June 9 debut, came back globally on July 1 once the controls were lifted on June 30. Gemini 3.5 Pro has missed three deadlines and still has no release date. The names to watch now are OpenAI's Astra, announced on August 1 with no date attached, and Gemini 4, which Google says is in training.

GPT-5.6: shipped July 9, rumors scored

In June this was the murkiest story of the three. Now that it has played out, the record is worth keeping, because the same pattern will repeat with the next launch.

What the reporting got right:

• The government gating was real. Axios reported on June 25, 2026 that the US government had asked OpenAI to limit the initial release, and that is how it went: a limited preview opened in late June for a small set of government-vetted organizations while the Department of Commerce reviewed a broader launch.

• The "couple of weeks" timeline in Altman's internal memo was close. Public release came on July 9, 2026, across ChatGPT, the API, and Codex.

• The Codex backend signal held up. A routing reference to gpt-5.6 was reportedly found in OpenAI's Codex backend in June, suggesting canary testing, and the model shipped publicly on July 9.

What the rumor mill got wrong:

• The rumored 1.5M-token context window never materialized. GPT-5.6 Sol lists a 1.05M-token window with up to 128K output tokens, the same ceiling GPT-5.5 had.

• Polymarket's roughly 89% odds pointed at a public release by June 30; it came nine days late, on July 9. The market priced the sentiment well and the date badly.

• Everyone talked about "GPT-5.6" as one model. It shipped as three: Sol as the flagship, Terra as the mid-tier everyday model, and Luna as the fast, low-cost tier.

The bigger consumer story landed a month later. On August 6, 2026, OpenAI announced "unlimited" free messages in ChatGPT, running on GPT-5.6 Luna as the default free model, with ads rolling out and caps still in place on files, images, voice, and tools. If you are weighing that free lane against the $8 per month ChatGPT Go tier, our look at whether ChatGPT Go is worth it breaks down what each one actually includes.

Claude Mythos: pulled, restored, repriced

The naming has needed caution since the June rumor cycle, and it still does. Anthropic never sold a product called "Mythos 1"; its own wording is "Mythos-class models," and the models that exist are Claude Fable 5 (the general release, with safety classifiers) and Claude Mythos 5 (the same underlying model with some safeguards lifted, restricted to Project Glasswing partners). The timeline has since gained two chapters:

• May 28, 2026 (announced): alongside Claude Opus 4.8, Anthropic said Mythos-class models would reach all customers "in the coming weeks."

• June 9, 2026 (shipped): Fable 5 and Mythos 5 went generally available on the Claude API, AWS Bedrock, Google Cloud, and Microsoft Foundry.

• June 12, 2026 (pulled): both were disabled for all customers under a reported US export-control directive, citing security concerns. It was described at the time as the first export-control order issued against a commercial LLM.

• June 30, 2026 (cleared): the government lifted the controls after a review conducted with Anthropic, per CNBC and Anthropic's own announcement.

• July 1, 2026 (restored): Fable 5 returned globally across Anthropic's platform. Mythos 5 remains limited to Project Glasswing partners.

Here is the part the comeback headlines undersold: getting Fable 5 back online did not put it back within reach. As of August 2026, Anthropic reserves Fable 5 access for its $100 per month Max tier, and the $20 Claude Pro plan gets the Opus and Sonnet families instead. A model can be restored and still move away from you.

Meanwhile Anthropic's regular cadence kept going. Claude Sonnet 5 shipped with introductory API pricing of $2 input and $10 output per million tokens through August 31, 2026 (then $3 and $15), and Claude Opus 5 arrived in late July at $5 and $25 per million tokens with a 1M-token context window. For how the previous round of flagships compared, see Claude Opus 4.8 vs GPT-5.5.

Gemini 3.5 Pro: the one that slipped

At Google I/O on May 19, 2026, Sundar Pichai said of Gemini 3.5 Pro: "Give us until next month to get it to you." That month came and went, and so did two more targets. By mid-July the model had missed three deadlines, with reporting pointing at coding performance and output reliability as the sticking points.

The miss got expensive. When Bloomberg reported the setback on July 16, 2026, Alphabet shares fell 4.4% that day. On July 21, Google shipped what it had instead: Gemini 3.6 Flash, a 1M-token-context model priced at $1.50 input and $7.50 output per million tokens, alongside Gemini 3.5 Flash-Lite and a government-gated security model, while confirming that Gemini 4 is in training. Worth noting: 3.6 Flash undercut the output price of the 3.5 Flash it replaced, a partial reversal of June's "cheap AI is ending" narrative. The wider repricing story is covered in our AI price war breakdown.

As of early-August reporting, Gemini 3.5 Pro still has no release date, and the specs that circulated in June (a 2M-token context window, a Deep Think mode, pricing near $15 and $60 per million tokens) remain unconfirmed. For where the three big ecosystems stand today, ChatGPT vs Claude vs Gemini covers current strengths.

The wave at a glance

One table, updated to mid-August 2026. Status first, headline second.

What is actually next

OpenAI's next act is called Astra. The company disclosed the name on August 1, 2026, folded into a blog post about math results, with no release date, no pricing, and no model card. What OpenAI has described is a system built around multiple agents that plan, test, and revise work together over hours or days, and the company has not said whether Astra becomes GPT-6 or another point release in the GPT-5 line. Treat every Astra spec you see between now and an official product page as a rumor; that rule scored well this summer.

There is already a complication. On August 7, 2026, OpenAI said it had slowed parts of Astra's development after internal evaluations left it unable to rule out critical cyber capabilities under its Preparedness Framework, the level at which a model could identify and carry out attacks against well-protected systems. Set that next to GPT-5.6's government-vetted preview, the Fable 5 export-control episode, and Google gating its security-focused Flash model, and the pattern is clear: security review is now a standing checkpoint on every lab's release calendar.

Two more items for the watchlist. Google has confirmed Gemini 4 is in training, with 3.5 Pro possibly arriving as a stopgap before it; neither has a date. GPT-6 remains unconfirmed in any form. And products get cut as fast as they get added: OpenAI shut down its Atlas browser on August 9, 2026, and our post on what Atlas users should do now covers that fallout.

Should you wait, or subscribe now?

Back in June it was already clear that betting your subscription on a single launch date is a bad bet. The summer tested that idea from every direction. If you had waited for GPT-5.6, you got it on July 9, more than a week after the prediction markets expected. If you had waited for Fable 5, it came back on July 1 and then priced itself into a $100 tier. If you had waited for Gemini 3.5 Pro, you are still waiting. Nobody who bet on one lab's calendar had a good summer, including the labs.

What you can control is being positioned to use whatever ships, whenever it ships. That is the case for a multi-model setup over a single-provider subscription. MultiChats carries GPT-5.6 Sol, Terra, and Luna, Claude Opus 5 and Sonnet 5, Gemini 3.6 Flash, Grok 4.3, and 25+ models in all under one plan, and you can switch models mid-conversation without losing the thread. When the next name on the watchlist actually ships and joins the lineup, you put it next to the model you were already using, in the same chat, and judge it on your own work instead of on a launch post.

If you are weighing where to land, the rundown of the best ChatGPT alternatives for 2026 and the case for keeping all AI models in one app both walk through the tradeoffs in more depth.

FAQ

When did GPT-5.6 come out?

OpenAI released GPT-5.6 publicly on July 9, 2026, in three tiers named Sol, Terra, and Luna, across ChatGPT, the API, and Codex. A limited preview had opened in late June for a small set of government-vetted organizations while the US Department of Commerce reviewed the release. Since August 6, GPT-5.6 Luna has also been the default model behind ChatGPT's unlimited free messages.

What is Claude Mythos, and can you use it?

Mythos is Anthropic's name for a model tier above the Opus family. Claude Mythos 5 shipped on June 9, 2026, was suspended under an export-control directive on June 12, and even after the controls were lifted on June 30 it remains restricted to Project Glasswing partners. Its general-release twin, Claude Fable 5, returned globally on July 1, and as of August 2026 Anthropic gates it behind the $100 per month Max tier rather than the $20 Pro plan.

Is Gemini 3.5 Pro released yet?

No. As of mid-August 2026 it has missed three deadlines since being promised "next month" at Google I/O on May 19, and Google has not committed to a new date. On July 21 Google shipped Gemini 3.6 Flash and Gemini 3.5 Flash-Lite instead and confirmed Gemini 4 is in training. The 2M-token context window and Deep Think mode attributed to 3.5 Pro remain unconfirmed.

What is OpenAI Astra?

Astra is OpenAI's announced next major model, disclosed on August 1, 2026 with no release date or pricing. OpenAI describes it as an agentic system in which multiple agents plan, test, and revise work together over long stretches. On August 7 the company said it had slowed parts of Astra's development after internal evaluations left it unable to rule out critical cyber capabilities under its Preparedness Framework. OpenAI has not said whether it will ship as GPT-6 or under another name.

Should I wait for the next model before subscribing?

The summer of 2026 is a strong argument against it. GPT-5.6 arrived weeks after the rumored dates, Gemini 3.5 Pro has not arrived at all, and the model that did ship on schedule, Claude Fable 5, was suspended within three days and later moved to a $100 tier. A multi-model plan like MultiChats sidesteps the timing problem: you get current frontier models now, including GPT-5.6 and Claude Opus 5, and new ones as they ship and join the lineup, with no second subscription and no migration.

The take

What June hinted at, the summer confirmed: the release calendar belongs to governments and safety teams as much as to the labs, and now the labs are pausing themselves, as OpenAI just did with Astra. Scoring the rumors showed a decent hit rate on whether things existed and a poor one on when and how they would ship. That is exactly the environment where tying yourself to a single provider's roadmap costs you. Be set up to use whatever wins, the moment it actually ships.

Want the next model available without another subscription decision? See MultiChats pricing and stop timing the release calendar.