Claude Opus 5 Price, Breaking Changes and Release Date

Claude Opus 5 is out. Anthropic released it on July 24, 2026 as claude-opus-5, at the same $5 / $25 per million tokens Opus 4.8 charged — after two weeks in which the model showed up in a code editor under a codename, vanished, and had its name leak through an error dialog. The leak trail turned out to be right about the name and wrong about the date.

Status — July 24, 2026 (UTC):

• Released? Yes, July 24, 2026. API ID claude-opus-5, a pinned snapshot with no date suffix.

• Price: $5 / $25 per million input/output tokens — unchanged from Opus 4.8, and half of Fable 5's $10 / $50.

• Specs: 1M-token context (default *and* maximum), 128k max output, effort ladder up to max, thinking on by default, knowledge cutoff May 2026.

• Where: Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, plus claude.ai, Claude Code and Cowork. It is the new default on Claude Max and the strongest model on Claude Pro.

• Two breaking changes before you swap the model ID: thinking is on by default, and thinking: {"type": "disabled"} now returns a 400 error at xhigh or max effort.

Claude Opus 5 API pricing and specs

Anthropic held the Opus tier at $5 / $25 while calling this release a "step change" over 4.8. At max effort Opus 5 lands within 0.5% of Fable 5's peak CursorBench 3.2 score at half the cost per task, so the case for paying Fable 5's $10 / $50 is now narrower.

Rates above are from Anthropic's pricing page; specs from What's new in Claude Opus 5.

Anthropic's published results are all relative to other models rather than absolute scores, and none has been independently replicated yet — read them as vendor-reported: Frontier-Bench v0.1, where Opus 5 "more than doubles Opus 4.8's performance at a lower cost per task"; ARC-AGI 3 at three times the next-best model; OSWorld 2.0 above Fable 5's best result at just over a third of the cost; Zapier's AutomationBench at roughly 1.5× the next-best pass rate for the same cost per task. On internal life-sciences evals it beats 4.8 by 10.2 points on organic chemistry and spectroscopy and 7.7 on protein tasks, and it scores 2.3 on Anthropic's misaligned-behavior audit, the lowest of the recent models. The Frontier-Bench figures come from Anthropic's own runs with mini-SWE-agent, averaged over five attempts per task.

Two limits are stated openly. Opus 5 trails Mythos 5 on cybersecurity and biology, and on OSS-Fuzz it finds vulnerabilities about as well as Mythos 5 but is "far behind" on exploit development. Mythos 5 stays invitation-only through Project Glasswing, so for most teams this is a note about what Opus 5 is not aimed at.

Customer figures relayed in the announcement, with no published methodology: Box reports Opus 5 beating 4.8 by 8% overall (11% on data analysis, 17% on due diligence), and an unnamed legal partner is credited with 26% fewer tokens on average at max reasoning.

Our own smoke test — availability, not capability. At 19:20 UTC on July 24, 2026, via Claude Code 2.1.219 headless (claude -p "<prompt>" --model <id> --output-format json, default effort high, one run each), we asked both models how many times the digit 7 appears in 1–1000. Both answered correctly (300): Opus 5 in 6.76s with 261 output tokens, claude-opus-4-8 in 7.37s with 239. Cost is omitted because cache state differed between runs. One prompt on launch day — it shows the new ID resolves and answers, not that it is faster.

Confirmed vs. rumor: the scorecard

How the July claims held up:

What to use right now (and how to access Opus 5 on day one)

With Opus 5 out at Opus 4.8's price, the default picks changed:

• Opus 5 ($5 / $25) for complex agentic coding, long-horizon tasks and enterprise work — Anthropic's own recommended starting point.

• Fable 5 ($10 / $50) when you need the capability ceiling and the CursorBench gap is worth double the token cost.

• Sonnet 5 ($3 / $15, introductory $2 / $10 through August 31) for high-volume, cost-sensitive work.

• Opus 4.8 only if you have a reason to pin it — a validated eval suite, or an integration that needs thinking disabled at high effort.

Switching is a model-ID change plus two checks:

• Thinking is on by default. Because max_tokens caps thinking plus response text together, workloads that ran 4.8 without thinking need their limits revisited.

• Disabling thinking is now restricted. thinking: {"type": "disabled"} is only accepted at effort high or below; pair it with xhigh or max and the request 400s.

In practice that means the invalid combination now looks like this:

// 400 error on Opus 5: thinking cannot be disabled above effort "high"
{ "model": "claude-opus-5", "max_tokens": 64000,
  "output_config": { "effort": "max" },
  "thinking": { "type": "disabled" } }

// Valid: keep max effort and drop the thinking field
{ "model": "claude-opus-5", "max_tokens": 64000,
  "output_config": { "effort": "max" } }

// Also valid: keep thinking disabled and lower effort to "high" or below
{ "model": "claude-opus-5", "max_tokens": 8000,
  "output_config": { "effort": "high" },
  "thinking": { "type": "disabled" } }

Anthropic's migration notes also flag two behaviour changes that appear without any code change: responses and written deliverables run longer, and the model verifies its own work unprompted — so carried-over instructions like "include a final verification step" now cause over-verification and should come out of your prompts.

On access: the Claude API is version-addressed, so a new model arrives as a new ID rather than a change to an existing one, and Anthropic, the clouds and Anthropic-compatible gateways each list it on their own schedule — a new ID can be live on one endpoint before another. Gateways set their own rates for the same IDs: AIReiter lists Opus 4.8 at $1.56 / $7.76 per MTok against the official $5 / $25, roughly 69% below, though as of July 24 its Claude list stops at Fable 5, Sonnet 5, Opus 4.8 and 4.7 — worth knowing if you stay pinned to 4.8 while validating Opus 5, not a route to Opus 5 itself yet.

When could Claude Opus 5 land? What the release cadence says

Opus 4.6 shipped February 5, 2026, Opus 4.7 followed 70 days later, Opus 4.8 another 42 days on — and Opus 5 arrived 57 days after that, inside the range the previous gaps set. The weekday theory did not survive: 4.6, 4.7 and 4.8 all landed on Thursdays, which sent speculation to July 23 and then July 30, and Opus 5 shipped on a Friday. So the interval range narrowed the window to about a fortnight, and a three-release weekday streak still called the wrong day. Anthropic publishes nothing in advance, which leaves the models documentation and its news page as the places a launch shows up first.

What actually leaked: the Honeycomb sighting in Cursor

For the record, the trail that ran ahead of the announcement. Around July 8 a listing labelled "Honeycomb EAP 1M Extra High" appeared briefly in Cursor, the AI code editor — developer @chetaslua posted screenshots, The New Stack covered it, then the entry disappeared. Three details in it match the shipped model: the 1M-token context window, the "extra high effort" setting (Opus 5 went one tier further with max), and safety fallbacks, which the launch announcement later described as flagged requests falling back to Opus 4.8 on claude.ai, Claude Code and Cowork, opt-in on the API.

For two weeks that screenshot was the entire artifact. On July 24, hours before the announcement, screenshots on X showed a Cursor error dialog naming the model outright: "The model 'claude-opus-5-thinking-high' requires Max Mode to be enabled" — the first appearance of the literal string claude-opus-5 in a shipping product. Deployment trackers had called the rollout a day early, with @M1Astra reporting it "rolling out across providers" on July 23 while Anthropic said nothing until the 24th.

!Cursor error dialog titled Max Mode Required, saying the model claude-opus-5-thinking-high requires Max Mode to be enabled

Is Honeycomb actually "Claude Opus 5"?

Functionally the question is closed: the product is Claude Opus 5, the API ID is claude-opus-5. Anthropic simply never connected the codename to it — no launch material mentions Honeycomb, so the link stays an inference, if a well-supported one. The naming worry behind the speculation, that Anthropic had dropped the Opus 5 name for the Fable and Mythos tiers introduced on June 9, did not hold: the lineup now runs Opus 5 as the agentic-coding default, Fable 5 as the capability ceiling, Sonnet 5 for speed and cost, and Mythos 5 as the invitation-only security tier.

It is the second time the pattern has played out — a claude-sonnet-5 identifier (codename Fennec) appeared in Google Vertex AI logs before Sonnet 5 — which makes pre-launch strings in third-party tools a reliable signal of *what* is coming and a poor one for *when*.

FAQ

Is Claude Opus 5 out yet?

Yes. Anthropic released Claude Opus 5 on July 24, 2026, available immediately as claude-opus-5 on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, claude.ai, Claude Code and Cowork.

How much does Claude Opus 5 cost?

$5 per million input tokens and $25 per million output tokens — the same as Opus 4.8. Batch API halves that to $2.50 / $12.50, cache hits read at $0.50 per MTok, and fast mode (a research preview on the Claude API only) runs at roughly 2.5× the default speed for $10 / $50.

Is Claude Opus 5 cheaper than Opus 4.8?

No — the per-token rates are identical. The "half the price" line circulating with the launch compares Opus 5 to Fable 5 at $10 / $50, and Anthropic's per-task cost comparisons come from benchmark runs where Opus 5 needed fewer tokens or turns to finish the same work. On a like-for-like token basis, Opus 5 and Opus 4.8 bill the same.

Is Claude Opus 5 better than Fable 5?

It depends on the axis. Fable 5 is still Anthropic's most capable released model, but Opus 5 came within 0.5% of Fable 5's peak CursorBench 3.2 score at max effort for half the cost per task, and beat Fable 5's best OSWorld 2.0 result at about a third of the cost. Fable 5 costs $10 / $50 against Opus 5's $5 / $25.

Is Honeycomb the same as Claude Opus 5?

Almost certainly, though Anthropic never said so. The Honeycomb listing that appeared in Cursor in early July matched Opus 5's 1M context window, xhigh effort setting and Opus 4.8 safety fallback, but no launch material mentions the codename.

Will Claude Opus 5 be free?

Not on the API. On claude.ai it is the default model for Claude Max subscribers and the strongest model available on Claude Pro; free-tier access to Opus-class models has always been capped.

Should I upgrade from Opus 4.8 to Opus 5?

Probably, for agentic and coding work: it is the same price with better vendor-published results, and migration is a model-ID change. Since those results have no independent replication yet, run your existing eval suite on Opus 5 before switching production traffic, and check two things: thinking is on by default and counts against max_tokens, and disabling thinking now fails at xhigh or max effort.

Related reading

• Claude Mythos vs Opus 4.8

• Fable 5 vs GPT-5.5