Foundations · Layer I / Broad

The new Claude lineup and what it breaks

Claude's July 2026 lineup is four models: Fable 5, Glasswing-gated Mythos 5, Sonnet 5 at intro pricing, and Opus 4.8, plus the three migration changes that bite.

Nº 013 · Vol. I·5 min read· Updated July 2026

“this quarter’s Claude lineup already replaced last quarter’s”

Moving a working prompt onto a current Claude model can hand you a 400 error, a 30% bigger token bill, or a retirement deadline you did not know about. This guide is a snapshot from one date (mid-July 2026, per the official release notes) of what changed and what bites.

Four models matter right now. Claude Fable 5 (claude-fable-5) is the most capable widely released model, launched June 9, 2026 alongside Claude Mythos 5 (claude-mythos-5), which is restricted to Project Glasswing participants and not something you can build on. Claude Sonnet 5 (claude-sonnet-5) launched June 30 at introductory pricing of $2/$10 per MTok (dollars per million input tokens, then output tokens), rising to $3/$15 after August 31, 2026. Claude Opus 4.8 (claude-opus-4-8) launched May 28 and stays widely available.

Model	ID	Status
Claude Fable 5	`claude-fable-5`	Most capable widely released model, launched June 9, 2026
Claude Mythos 5	`claude-mythos-5`	Project Glasswing participants only
Claude Sonnet 5	`claude-sonnet-5`	$2/$10 per MTok intro pricing until August 31, 2026, then $3/$15
Claude Opus 4.8	`claude-opus-4-8`	Widely available, launched May 28, 2026

The shared spec is the easy part to remember: all four default to a 1M token context window and 128k max output tokens. Exact pricing for Fable 5, Mythos 5, and Opus 4.8 is not published on the release-notes page, so treat any number you have for them as unconfirmed until you check the model reference.

02Three changes that return errors

The migration pain is concentrated in three behavior shifts, and two of them fail loudly. First, thinking is always on and adaptive. On Fable 5 and Mythos 5 there is no other mode. Manual extended thinking (thinking type enabled with budget_tokens) returns a 400 error, and so does thinking type disabled. If your code sets a thinking budget, it stops working. Use the effort parameter or max_tokens instead.

Second, non-default sampling is gone. Passing a non-default temperature, top_p, or top_k returns a 400 error on Sonnet 5 and Opus 4.8. Any prompt that pinned temperature to zero for determinism needs that line removed.

Third, and quietest, the tokenizer. Fable 5, Mythos 5, and Sonnet 5 use the tokenizer introduced with Opus 4.7, which produces roughly 30% more tokens for the same text than older models.

“Nothing errors. Your prompts just cost about a third more and hit output limits sooner than the old count predicted.” Re-measure before you migrate: the token counting API accepts claude-fable-5 to count under the new scheme. And once the bigger bill lands, Thinking on a budget is the guide to cutting reasoning-token spend without cutting depth.

03Refusals and deadlines

Fable 5 runs safety classifiers before and during generation. A declined request comes back with stop_reason refusal, and since June 2, 2026 a refusal that generated no output is not billed. The stop_details.category field tells you which classifier fired: the familiar cyber and bio, plus a newer reasoning_extraction, which blocks requests aimed at reverse-engineering or duplicating the model’s output. It needs no beta header. If you audit skills, watch for show-your-thinking language that could trip it, a hazard Guardrails for hours-long agent runs covers in depth. An opt-in fallbacks beta can re-run a refused request on another model at that model’s rates, though not on the Message Batches API.

Then the deadlines. Claude Sonnet 4 and Claude Opus 4 were retired June 15, 2026 and now return errors. Claude Opus 4.1 is being phased out, with retirement scheduled for August 5, 2026.

“Any code still pointing at a retired model ID stops working on its retirement date.”

And Sonnet 5’s cheap intro pricing ends August 31, so cost math done today expires in a few weeks. Treat previews the same way: Mythos 5 is invitation-only, and fast mode on Opus 4.8 is a research preview. Neither belongs in a plan you need to rely on.

FAQ

You move a working prompt to Claude Sonnet 5 and it starts returning 400 errors. What three changes should you check first?

Adaptive thinking is on by default and manual extended thinking (thinking type "enabled" with budget_tokens) now returns 400; non-default temperature, top_p, or top_k values return 400; and the model uses the new Opus-4.7 tokenizer, so any hard token budget you set against the old count may now overflow. The first two throw errors directly; the third is a silent cost and length shift.

A request to Claude Fable 5 comes back with stop_reason "refusal" and no output. What does that cost you, and what does stop_details.category tell you?

It costs nothing. Since June 2, 2026 the API does not bill refusals that generated no output. The category names why a safety classifier declined: cyber, bio, or the newer reasoning_extraction, which fires on requests aimed at reverse-engineering or duplicating the model, and needs no beta header to appear.

Why does the release-notes fact sheet warn you to treat parts of the lineup as still changing?

Because several dates are already fixed and several features are previews. Opus 4.1 retires August 5, 2026 and Sonnet 5 intro pricing ends August 31, 2026, so anything built against them has a deadline. Mythos 5 is Glasswing-only, and fast mode on Opus 4.8 is a research preview, so neither is something to depend on in production.

Sources

Original work by the editor, no video source.