Claude in Q2 2026: Opus 4.8, Fable 5, Sonnet 5, and a Two-Week Blackout

Q1 2026 was dense. Q2 was stranger.

Anthropic shipped four frontier models in three months, held its first developer conference, and then watched two of those models get pulled offline worldwide by a US government directive for eighteen days. By June 30 the lineup had been rebuilt from the ground up: the Claude 5 family replaced almost everything that existed on April 1.

This article covers every significant release from April through June 2026, including the API changes that will break code written against Opus 4.6. If you want the previous three months, the Q1 2026 releases post covers January through March.

April: Opus 4.7 and the Tooling Buildout

April has its own detailed walkthrough, so the short version here. Claude Opus 4.7 (claude-opus-4-7) promoted the 1M-token context window out of beta and became the default for new Max-tier Claude Code sessions. Alongside it: Claude Code Remote Control, a hardened Agent SDK, Fast mode on Opus, and first-class plugins and skills.

Two Opus 4.7 changes matter for everything that follows, because they carried forward into every later model:

• Manual extended thinking was removed. thinking: {type: "enabled", budget_tokens: N} returns a 400. Adaptive thinking plus an effort level replaced it.
• Sampling parameters were removed. temperature, top_p, and top_k return a 400.

If your code still sets either, it stops working the moment you move past Opus 4.6. That is the single most common migration failure of the quarter.

Code with Claude: San Francisco, May 6

Anthropic’s first developer conference ran in San Francisco on May 6, with sessions published to YouTube afterward. Coverage spanned Claude Code, the Claude Developer Platform, and partner deployments at GitHub, Vercel, Datadog, and Bun.

Three announcements came out of it (InfoQ’s writeup is the best summary):

• Managed Agents. Server-managed stateful agents where Anthropic runs the agent loop and hosts a per-session container for tool execution. You define a persisted, versioned agent config once, then start sessions that reference it.
• Proactive Workflows. Agents that initiate work rather than waiting to be prompted.
• The Capability Curve. Anthropic’s framing for how model capability translates into task horizon.

Managed Agents is the one with the largest practical footprint. It is a genuinely different surface from the Messages API: agents, sessions, environments, and vaults are all persisted objects with their own endpoints.

Claude Opus 4.8: May 28

Opus 4.8 (claude-opus-4-8) arrived at the same price as 4.7: $5 per million input tokens, $25 per million output. Same request surface too, which made it the rare model upgrade that is genuinely just a model-ID swap.

Shipped alongside it:

• Effort control on claude.ai. End users, not just API callers, can now dial how much work Claude puts into a task.
• Dynamic workflows in Claude Code. For tackling very large-scale problems.
• Cheaper fast mode. Fast mode runs Opus at roughly 2.5x the output speed, and on 4.8 it landed at a third of the cost it carried on previous models.

Behavioral shifts worth knowing: 4.8 narrates more than 4.7 during long agentic sessions, writes in a warmer and less hedged voice, and is more conservative about reaching for search, subagents, and file-based memory. That last one is the trap. If you migrated from 4.7 and your tool-call rate dropped, the fix is explicit triggering guidance in the tool’s own description field, not more emphasis in the system prompt:

Call this when the user asks about current prices or recent events.

Prescriptive “call this when” descriptions measurably outperform descriptions that only state what the tool does.

Claude Fable 5 and Mythos 5: June 9

On June 9 Anthropic released Claude Fable 5 (claude-fable-5), its most capable widely released model, alongside Claude Mythos 5 (claude-mythos-5), the same model available only through Project Glasswing. State of the art on Cognition’s FrontierCode evaluation and on vision.

Specs: 1M context window as both default and maximum, 128K max output, $10 per million input tokens and $50 per million output. That pricing sits above the Opus tier, which is the point: Fable 5 is not the default Opus upgrade path, it is the model you reach for on the hardest long-horizon work.

The API surface is different enough to matter:

# Fable 5: thinking is always on. Do not configure it.
client.messages.create(
    model="claude-fable-5",
    max_tokens=16000,
    output_config={"effort": "high"},  # low | medium | high | xhigh | max
    messages=[...],
)

• Thinking is always on. Omit the thinking parameter entirely. Both {type: "disabled"} and {type: "enabled", budget_tokens: N} return a 400.
• The raw chain of thought is never returned. You get thinking blocks whose text is a summary when you set display: "summarized", and empty otherwise.
• No assistant prefill. Use structured outputs via output_config.format instead.
• 30-day data retention is required. Organizations configured for zero data retention get a 400 invalid_request_error on every request, with a perfectly valid payload. If a migration suddenly fails for no visible reason, check the retention setting before debugging the request body.
• Safety classifiers can decline a request. You get HTTP 200 with stop_reason: "refusal", not an error. Code that reads response.content[0] unconditionally breaks.

Handle the refusal path. Check stop_reason before reading content. The classifiers target research biology and most cybersecurity content, and benign adjacent work (security tooling, life-sciences tasks) trips them often enough that a fallback is worth wiring in from day one.

The Suspension: June 12 to June 30

Three days after launch, both models went dark worldwide.

US Commerce Secretary Howard Lutnick sent Anthropic CEO Dario Amodei a letter directing the company to place Fable 5 and Mythos 5 under export controls, prohibiting access “by any foreign national, whether inside or outside the United States.” The directive was delivered June 12; Anthropic received formal notification at 5:21 p.m. ET on June 13, and both models were offline for all users globally within hours.

The trigger was a prompting technique, discovered by Amazon researchers, that bypassed Fable 5’s safeguards around identifying software vulnerabilities. Anthropic published a statement on the directive at the time.

On June 30 the controls were lifted. Anthropic’s redeployment announcement confirmed Fable 5 returning globally on July 1 across the Claude Platform, claude.ai, Claude Code, and Claude Cowork.

The practical lesson is about dependency risk, not politics. An eighteen-day outage on a frontier model is a product-level event. If a workload depends on a single model ID with no configured fallback, that workload had eighteen days of downtime. Server-side fallbacks and a tested second-choice model are cheap insurance.

Claude Sonnet 5: June 30

Sonnet 5 (claude-sonnet-5) replaced Sonnet 4.6 as the default model for Free and Pro plans, and is available across Max, Team, and Enterprise. It reaches what was previously Opus-tier quality on a lot of coding and agentic work.

Pricing: $3 per million input, $15 per million output, with an introductory rate of $2 and $10 running through August 31, 2026.

What changed under the hood:

Change	Detail
Adaptive thinking	On by default. Omitting thinking now runs adaptive, where 4.6 ran thinking-off
Sampling parameters	Non-default temperature / top_p / top_k rejected
Extended thinking	budget_tokens fully removed, the 4.6 escape hatch is gone
Tokenizer	New. Roughly 30% more tokens for the same text
Effort	First Sonnet with the full low through max ladder, including xhigh
Vision	First Sonnet with high-resolution images, 2576px on the long edge
thinking.display	Defaults to "omitted", where 4.6 defaulted to "summarized"

The tokenizer change is the one that bites quietly. Per-token pricing did not change, but the same text now counts about 30% higher, so a max_tokens value tuned against Sonnet 4.6 can truncate equivalent output. Re-run count_tokens against the new model rather than applying a multiplier:

resp = client.messages.count_tokens(
    model="claude-sonnet-5",
    messages=[{"role": "user", "content": open("prompt.md").read()}],
)
print(resp.input_tokens)

Combined with adaptive thinking now being on by default, a caller who never set thinking gets both more thinking spend and empty-text thinking blocks in the stream. If you render reasoning to users, set display explicitly.

Claude Science and Claude Desktop for Linux

Two more launches landed on June 30.

Claude Science (beta) is an AI workbench for researchers, integrating tooling for genomics, single-cell analysis, proteomics, structural biology, and cheminformatics. Available to Pro, Max, Team, and Enterprise subscribers on macOS and Linux.

Claude Desktop for Linux (beta) supports Ubuntu and Debian. It is the first official Linux desktop client, which closes a gap that had been open since the desktop app launched.

The Lineup After Q2

Model	Model ID	Context	Input / 1M	Output / 1M
Claude Fable 5	claude-fable-5	1M	$10	$50
Claude Opus 4.8	claude-opus-4-8	1M	$5	$25
Claude Opus 4.7	claude-opus-4-7	1M	$5	$25
Claude Sonnet 5	claude-sonnet-5	1M	$3	$15
Claude Haiku 4.5	claude-haiku-4-5	200K	$1	$5

Picking one: Fable 5 only when someone explicitly chose it, because the pricing is above Opus tier. Opus 4.8 for demanding agentic and coding work. Sonnet 5 as the default for most production workloads, since it now covers a lot of what needed Opus in Q1. Haiku 4.5 for mechanical, no-judgment tasks.

The Migration Checklist

If you are moving code from an Opus 4.6 era baseline, these are the changes that cause a 400 rather than a quality shift:

• Replace thinking: {type: "enabled", budget_tokens: N} with thinking: {type: "adaptive"} plus output_config.effort
• Strip temperature, top_p, and top_k
• Remove last-assistant-turn prefills, use output_config.format
• Move output_format into output_config.format
• Stream anything over roughly 16K max_tokens
• On Fable 5, remove the thinking parameter entirely and confirm 30-day data retention
• Add stop_reason == "refusal" handling before reading response.content

And the quality-level adjustments that will not error but will change behavior:

• Re-baseline token counts with count_tokens on the new model, especially on Sonnet 5
• Re-tune effort per route rather than inheriting the old default
• Set thinking.display: "summarized" if you surface reasoning to users
• Add explicit “call this when” triggering language to tool descriptions
• Dial back CRITICAL: and MUST language, which now overtriggers

Summary: Q2 2026 by the Numbers

• 4 frontier model releases (Opus 4.7, Opus 4.8, Fable 5 with Mythos 5, Sonnet 5)
• 18 days Fable 5 and Mythos 5 were offline worldwide
• 1 developer conference, with Managed Agents as its largest announcement
• 30% more tokens for the same text under the Sonnet 5 tokenizer
• 2 parameters that now hard-fail on current models (budget_tokens, sampling)

The through-line for the quarter is consolidation. Q1 built the agent platform: teams, scheduling, memory, computer use. Q2 rebuilt the models underneath it and made the API surface consistent across the whole lineup. The awkward part is that “consistent” meant removing things, so the migration is real work rather than a version bump.

Official sources:

• Anthropic News
• Claude Models Overview
• Model Migration Guide
• Claude Code Changelog
• Claude Code GitHub Releases