Claude Opus 5: Fable 5 intelligence at half the price, with a dial for how hard it thinks

Last updated: July 26, 2026

TL;DR

Claude Opus 5 is Anthropic's new flagship Opus model, released July 24, 2026. It costs $5 per million input tokens and $25 per million output tokens – the same as Opus 4.8, and half of Claude Fable 5 – while landing within 0.5% of Fable 5's peak coding score. Its API model string is claude-opus-5.

What is Claude Opus 5?

Claude Opus 5 is the newest model in Anthropic's Opus tier, announced on July 24, 2026. Anthropic describes it as a "thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price." It is the new default model on Claude Max and the strongest model available on Claude Pro.

The pitch is not a new capability tier. Fable 5 still sits at the top. The pitch is that the gap between the top tier and the tier most people actually pay for got very small, and the price did not move.

What Claude Opus 5 costs

Opus 5 ships at Opus 4.8's exact price. That matters more than the benchmark numbers, because it means the upgrade is a model-string swap with no budget conversation attached.

Two footnotes on cost. Fast mode runs Opus 5 at roughly 2.5 times the default output speed for double the price ($10 / $50), and it is available on the Claude API and Claude Code usage credits only – not on Amazon Bedrock, Google Cloud, or Microsoft Foundry. And Opus 5 draws from its own rate-limit bucket, separate from the shared pool that Opus 4.8, 4.7, 4.6, and 4.5 share. Moving traffic over does not free up headroom on the old bucket or inherit it, so check your tier limits before you shift volume.

The benchmarks that matter

Anthropic's own numbers, with the cost comparison attached where they published one:

The pattern across those rows is cost-per-result, not raw score. Opus 5 does not beat Fable 5 on coding – it gets within half a percent of it for half the money. On computer use it does beat Fable 5, at roughly a third of the cost. For anyone running agents at volume, that ratio is the story.

Anthropic is also explicit about where Opus 5 does not lead. It says the model "does not advance the frontier in risky, dual-use capabilities," and it remains behind Mythos 5 on biology research and cybersecurity exploitation. On OSS-Fuzz it comes close to Mythos 5 at identifying software vulnerabilities but is considerably less successful at developing exploits.

The effort dial, and what the coverage got slightly wrong

Most of the launch coverage described the effort setting as a new feature. It is not new to the API – output_config.effort has shipped on Opus-tier models since 4.5. Two things about it are genuinely new with Opus 5, and both are worth knowing.

First, Opus 5 supports the full ladder: low, medium, high, xhigh, and max, with high as the default. Second, and more useful, the low end of that ladder holds up unusually well on this model.

Anthropic frames the dial as something "customers can use to optimize for intelligence or conserve tokens for faster and cheaper results." The practical version: start at xhigh for coding and agentic runs, then sweep downward against your own evals. Effort defaults carried over from a previous model are usually wrong here, and low and medium on Opus 5 frequently outperform xhigh on the model you are migrating from.

One thing the dial does not do: shorten Opus 5's visible responses. Lowering effort moves how much the model thinks, not how much it writes. Response length is a prompting problem, covered below.

The instruction that used to help and now hurts

Standard prompting advice for years has been to ask the model to check its own work. On Opus 5, that advice is actively counterproductive.

Anthropic describes Opus 5 as "much stronger at verifying its work and iterating carefully until it succeeds." It does this without being asked. Instructions that tell it to verify – "include a final verification step," "double-check your answer before responding," "use a subagent to verify" – now push it into over-verification, burning tokens and time on work it was already doing.

Anthropic's own migration guide is unusually direct about the fix: remove those instructions, and doing so reduces over-verification with no capability regression. This is a delete, not a rewrite. The same applies to harness-level scaffolding – separate verification passes you built for an earlier model are likely redundant now.

Three other behaviors flip direction from Opus 4.8, and each has a short prompt fix:

1. It writes longer. Both conversational responses and files it saves to disk. A brief conciseness instruction cut user-facing response length by about 20% in Anthropic's testing. Add one, and add a separate length instruction for written deliverables if your product ships Claude-authored documents.
2. It delegates to subagents more freely. Opus 4.8 under-reached for subagents and needed prompting to delegate. Opus 5 reaches for them readily, and each one multiplies cost and latency. If you added "delegate more" guidance for 4.8, take it out and add a cap instead.
3. It can expand the scope of a task. It sometimes adds steps you did not request. An explicit scope-discipline instruction – deliver what was asked at the scope intended, flag a better approach in a sentence but keep going – reduced this to nearly zero in testing.

None of these are bugs. They are a more autonomous model behaving autonomously, and they are all steerable with two or three sentences of prompt.

What changes if you build on the Claude Opus 5 API

Swapping the model string gets you most of the way, but two changes will bite if you skip them:

1. Thinking is on by default. On Opus 4.8, omitting the thinking parameter meant no thinking. On Opus 5, omitting it runs adaptive thinking. Since max_tokens caps thinking plus response text together, a route that sized max_tokens tightly around its answer can now truncate mid-response. Revisit max_tokens on every call site that never set thinking.
2. Disabling thinking is capped at high effort. Sending thinking: {type: "disabled"} alongside xhigh or max returns a 400. The check runs per request, so audit every call site, not just the first one. If a route needs low latency, medium effort with thinking on usually beats keeping thinking off.

Three smaller changes are worth picking up:

• The prompt-cache minimum drops to 512 tokens, down from 1,024 on Opus 4.8. Prompts you previously wrote off as too short to cache now create entries with no code change.
• Mid-conversation tool changes (beta) let you add or remove tools between turns without invalidating the prompt cache. Previously any edit to the tool set re-billed the entire prefix.
• Automatic fallbacks (beta) route a request declined by a safety classifier to another model instead of returning the refusal to you. Anthropic recommends the automatic mode over pinning a specific fallback model, because the right substitute depends on why the request was declined.

That last one connects to a stop reason worth handling explicitly. Opus 5 ships with elevated cybersecurity safeguards, and a declined request returns a normal HTTP 200 with stop_reason: "refusal" – not an error. Code that reads the first content block without checking stop_reason will break on a refusal.

Where you can use Claude Opus 5 today

Opus 4.8 stays available on every one of those surfaces, so nothing you have running today breaks.

The safety picture

Anthropic reports a behavioral audit score of 2.3 for Opus 5, the lowest rate of misaligned behavior among recent models, and calls it their "most aligned model to date."

The more practical detail for anyone who has hit a false-positive refusal: Opus 5's cyber classifiers are proportionally less restrictive than Fable 5's, intervening around 85% less often. If you do legitimate security work and found Fable 5 unusable because benign requests kept getting declined, Opus 5 is a meaningfully different experience.

Should you switch to Claude Opus 5?

Split the answer by what you actually do.

If you run a small business and use AI for drafting, support, research, and light automation: you are probably already on it without doing anything. Opus 5 is the new default on Claude Max and the top model on Claude Pro, so it arrived in your account on July 24. There is nothing to configure. If you are on a lower plan and paying per token through the API, Sonnet 5 still does routine work well for less.

If you build agents or automations on the API: this is the clearest upgrade Anthropic has shipped in a while, because the price did not move. Read the two breaking changes above, run one real workload, then sweep effort levels. The cost saving from finding the right effort level is usually larger than the cost difference between models.

If you were building on Fable 5 for capability reasons: test whether Opus 5 covers your workload. On coding it is within half a percent at half the cost, and on computer use it is ahead at a third of the cost. Fable 5 still wins on the hardest long-horizon reasoning, but the set of tasks that genuinely require it just got smaller.

If you use Claude Code daily: the effort sweep is the highest-value thing you can do this week. Fast mode is also worth trying on interactive work, where the 2.5x speed change is felt rather than measured.

Frequently asked questions

What is Claude Opus 5?

Claude Opus 5 is Anthropic's newest Opus-tier model, released July 24, 2026. Anthropic describes it as coming close to the frontier intelligence of Claude Fable 5 at half the price. It is the default model on Claude Max and the strongest model on Claude Pro. Its API model string is claude-opus-5.

How much does Claude Opus 5 cost?

Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens. That is identical to Claude Opus 4.8 and half the price of Claude Fable 5. Fast mode runs at about 2.5 times the default speed for double the price, on the Claude API and Claude Code only.

Is Claude Opus 5 better than Claude Fable 5?

It depends on the task. Opus 5 lands within 0.5% of Fable 5's peak coding score at half the cost, and beats Fable 5 on OSWorld 2.0 computer use at about a third of the cost. Fable 5 still leads on the hardest long-horizon reasoning.

What is the difference between Claude Opus 5 and Claude Opus 4.8?

Same price, higher capability. Opus 5 supports all five effort levels, halves the prompt-cache minimum to 512 tokens, and verifies its own work without being asked. Two breaking changes: thinking runs by default, and disabling thinking is only allowed at high effort or lower.

What is the Claude Opus 5 effort dial?

The effort setting controls how much the model thinks and spends per request, across five levels from low to max, defaulting to high. Anthropic frames it as a way to optimize for intelligence or conserve tokens. On Opus 5, low and medium perform unusually well on routine work.

Should I tell Claude Opus 5 to double-check its work?

No. Opus 5 verifies its own work without being asked, and verification instructions push it into over-verification that costs tokens and time. Anthropic's migration guide recommends deleting those instructions outright, noting it reduces over-verification with no loss of capability.

Is Claude Opus 5 available in Claude Code?

Yes. Claude Opus 5 is available in Claude Code, Claude.ai, Claude Cowork, and the Claude API, plus Amazon Bedrock, Google Cloud, and Microsoft Foundry. Fast mode is available through Claude Code usage credits and the Claude API only.

Do I need to change my code to use Claude Opus 5?

Usually just the model string, plus two checks. Thinking now runs by default, so any route that sized max_tokens tightly may truncate. And thinking: {type: "disabled"} combined with xhigh or max effort returns a 400 error. Audit every call site, not just the first.

Is Claude Opus 5 safe for security work?

Opus 5's cyber classifiers are proportionally less restrictive than Fable 5's, intervening around 85% less often, so benign security work is far less likely to be declined. Anthropic says the model does not advance the frontier in risky, dual-use capabilities and stays behind Mythos 5 on exploit development.

Should small businesses switch to Claude Opus 5?

Most already have. Opus 5 became the default on Claude Max and the top model on Claude Pro on July 24, so no action is needed. If you call the API directly for routine drafting and support work, Claude Sonnet 5 still handles it well at a lower price.

Want the practical read on how each model release changes day-to-day work? Join 760+ builders in the WotAI community, or get the weekly rundown in your inbox at wotai.kit.com.