Claude Opus 5: Anthropic's Near-Frontier Model at Half the Price of Fable 5

On July 24, 2026, Anthropic shipped Claude Opus 5 — and within a day it did something rare for a "cheaper" model: it went to the top of the independent leaderboards. Anthropic pitched Opus 5 as reaching close to the frontier intelligence of Claude Fable 5 at half the price, and it launched the same day across Claude.ai, the Claude API, Claude Code, and Claude Cowork. Since then, Artificial Analysis has scored it as the #1 model on its Intelligence Index. This guide covers what Opus 5 actually is, what's independently verified versus vendor-reported, its new effort dial and safety posture, what it costs in practice, and who should switch.

Every figure below is labeled by source, because at this speed the story changes daily: what read "no independent score yet" on launch day is now "independently #1." Benchmarks and prices move — verify before you commit budget.

TL;DR. Claude Opus 5 is priced at $5 per 1M input tokens and $25 per 1M output — identical to Opus 4.8 and exactly half of Fable 5's $10 / $50. On the Artificial Analysis Intelligence Index it now leads at 61 (max effort), the top model of 170 evaluated, edging Fable 5 (60) and matching its class at half the cost. Anthropic's own launch benchmarks back this up: Frontier-Bench v0.1 43.3% (vs Fable 5's 33.7%), plus leads on Zapier's AutomationBench and the OSWorld 2.0 computer-use benchmark. Two things make it more than a price cut: an effort dial (low/medium/high/max) that trades cost for capability per request, and the strongest safety posture of any Opus so far. If you already build on Claude, upgrading is low-risk and doesn't raise your bill; if you're choosing between flagships, Opus 5 is now the value-and-quality leader — validate it on your own workload.

Key takeaways

• Now independently #1. Artificial Analysis lists Claude Opus 5 (max effort) at 61 on its Intelligence Index — the top-ranked model, ahead of Fable 5 (60).

• Half the price of Fable 5. $5 / $25 per 1M tokens, the same as Opus 4.8 and half of Fable 5's $10 / $50 — frontier-leading quality at a mid-tier price.

• New effort dial. Low/medium/high/max settings trade tokens (cost) for capability; "max" is what tops the leaderboard, cheaper settings still beat most rivals.

• Built for agents and computer use. Anthropic emphasizes finishing multi-step tasks (AutomationBench) and driving a real UI (OSWorld 2.0), not just single answers.

• Safest Opus yet. Anthropic calls it "the most aligned Opus model," expecting its cyber classifiers to intervene about 85% less often than Fable 5 — fewer false refusals on legitimate work.

• Broad, immediate availability. Default model on Claude Max, strongest on Claude Pro, and live on the API, Claude Code, and Claude Cowork from day one.

A note on reading this brief: everything attributed to Anthropic is a vendor claim from launch materials; independent figures (the Intelligence Index) are attributed to Artificial Analysis. Where a number isn't yet published (a separate SWE-bench Verified score), we say so rather than guess.

What Claude Opus 5 actually is

Opus 5 is the newest member of Anthropic's Opus line — the "heavy" tier above Sonnet and Haiku. It slots in above Opus 4.8 and, on the independent Intelligence Index, has now edged ahead of Fable 5, Anthropic's June-2026 frontier flagship. What makes it notable is not a modest capability bump but the combination: near-top-of-the-board intelligence at a workhorse price. Anthropic describes it as a "thoughtful and proactive" model aimed at coding, agentic workflows, and everyday enterprise office work.

Concretely, Opus 5 costs $5 per 1M input tokens and $25 per 1M output tokens. That is the same as Opus 4.8, and half of Fable 5's $10 / $50. It is the default on Claude Max and the strongest option on Claude Pro, and it is available through the Claude API, Claude Code (for coding), and Claude Cowork (for collaborative agent work) from launch day. The context window is Opus-family 1M-class.

What's actually new

The clearest launch-day signal was Anthropic's Frontier-Bench v0.1, a hard reasoning-and-tasks suite: Opus 5 scores 43.3%, versus 33.7% for Fable 5 and 18.7% for Opus 4.8. On that benchmark, Anthropic's cheaper new model outscores its own pricier flagship — a claim that has since been echoed by independent scoring.

Two more vendor benchmarks fill out the "built for real work" story. On Zapier's AutomationBench, which measures whether a model carries a business task through to completion rather than merely starting it, Anthropic says Opus 5's pass rate is roughly 1.5 times the next-closest model at matching cost, and that even its cheapest effort setting beats every competitor's best result. On OSWorld 2.0, a computer-use benchmark, Anthropic reports Opus 5 outperforming rivals at every price point and clearing Fable 5's peak score using about a third of the budget. Opus 5 also improves on Opus 4.8 across every life-sciences evaluation Anthropic tracks, with the biggest jump — over 10 percentage points — on organic-chemistry tasks like reading molecular structure from spectroscopy data.

The effort dial: paying only for the capability you need

The headline feature — and the one most reviewers zero in on — is the effort setting. Opus 5 lets you choose how much effort the model spends on a request: low, medium, high, or max. Higher effort means more internal reasoning tokens (so higher cost and latency) and higher quality; lower effort conserves tokens for faster, cheaper answers. It is, in effect, a dial on your AI bill.

This is why Artificial Analysis lists several Opus 5 entries (medium, high, xhigh, max): each effort level is a different point on the cost-quality curve. The "max" setting is what currently tops the Intelligence Index at 61; the lower settings trade a little quality for meaningfully lower cost per task, and even they, per Anthropic, beat most competitors' best results. In practice this means one model can serve both your latency-sensitive, high-volume calls (low/medium) and your hardest reasoning jobs (high/max) — you tune per request instead of switching models.

The benchmark deep-dive: what these tests actually measure

It helps to know what each number means before trusting it. The Artificial Analysis Intelligence Index is a composite of many independent evaluations (reasoning, math, coding, knowledge) run by a third party; a score of 61 placing Opus 5 first out of 170 models is the most credible single signal here, precisely because Anthropic didn't run it. Frontier-Bench v0.1 (43.3% for Opus 5) is Anthropic's own hard set — a relative signal that Opus 5 clears more difficult problems than Fable 5 (33.7%) or Opus 4.8 (18.7%), not "43% correct" in a school sense.

AutomationBench and OSWorld 2.0 matter more than a single reasoning score if your use case is agents. AutomationBench rewards task completion — the difference between an agent that drafts an email and one that actually files the expense report — and OSWorld 2.0 rewards driving a real computer UI. Anthropic's claim that Opus 5 matches Fable 5's OSWorld peak at a third of the cost is, now that the Intelligence Index corroborates its general standing, a strong argument for moving agentic workloads to Opus 5. The one gap: there is no separate independently-audited SWE-bench Verified figure for Opus 5 yet (Fable 5 sits at 95.0%, Opus 4.8 at 88.6%), so for a pure coding-accuracy citation, wait for that number.

What it costs in practice

Per-token prices are abstract; per-task is what hits your invoice. Take a representative agent step of 20,000 input tokens and 3,000 output tokens. Opus 5 (and Opus 4.8): input 20k × $5/1M = $0.10, output 3k × $25/1M = $0.075, so about $0.175 per call. Fable 5 at $10 / $50: $0.20 input + $0.15 output = about $0.35 — roughly double. Over 100,000 such calls a month, that is about $17,500 on Opus 5 versus about $35,000 on Fable 5. Because the Intelligence Index now puts Opus 5 ahead of Fable 5, this is not "near-Fable-5 for less" — it is leading-class quality for roughly half the monthly bill.

The effort dial changes the math again. A high-volume support pipeline can run most calls at low or medium effort — cheaper than the numbers above — and reserve max effort for the handful of genuinely hard tickets. Anthropic's framing is that even the cheap settings beat rivals' best, so you rarely pay for capability you don't use. Model your own mix of effort levels rather than assuming every call runs at max.

How to access and use Claude Opus 5

Opus 5 is available immediately in five places: Claude.ai (where it is the default on the Max plan and the strongest option on Pro), the Claude API (model id in Anthropic's docs), Claude Code (for terminal and IDE coding), and Claude Cowork (for collaborative, multi-step agent work). API users set the effort level as a parameter, so the cost-capability trade-off is programmable per request. If you already call Opus 4.8, switching is a model-id change at the same price — the lowest-friction upgrade Anthropic has shipped in a while.

Who it's for: three scenarios

1. Agentic and computer-use pipelines

If your product runs multi-step agents or drives a browser or desktop, Opus 5's AutomationBench and OSWorld 2.0 results are aimed squarely at you — now backed by an independent #1 Intelligence Index. Run most steps at medium effort and escalate to high/max for the hard ones; pilot on your own agent traces before committing.

2. Coding teams watching the budget

Opus 5 is in Claude Code from day one at the Opus 4.8 price. With the Intelligence Index lead it is a strong default coder; the only missing citation is an independent SWE-bench Verified number (Fable 5 still owns that at 95.0%). For teams already on Claude, the upgrade is low-risk and free of a price increase.

3. Enterprises standardizing on one model

Broad availability, 1M-class context, the safest Opus alignment profile, and an effort dial that covers both cheap bulk work and hard reasoning make Opus 5 a plausible single default for mixed office + engineering work — at a cost that is far easier to defend than a frontier flagship.

Safety and alignment

Anthropic positions Opus 5 as its safest Opus release: "the most aligned Opus model," and "the least susceptible to being tricked into misuse." A practical consequence matters for real deployments: Anthropic expects Opus 5's cyber classifiers to intervene about 85% less often than Fable 5's. In plain terms, that should mean fewer false refusals on legitimate security, coding, or research work — a common frustration with heavily-guardrailed frontier models — while keeping genuine misuse blocked. As always, validate against your own content and compliance requirements rather than taking the posture on faith.

The Fable 5 export-control story is now in the system prompt

Fable 5’s launch was anything but smooth, and that context matters for anyone comparing Claude Opus 5 against it. Three days after Claude Fable 5 and Claude Mythos 5 shipped on June 9, 2026, Anthropic suspended both models to comply with a U.S. Commerce Department export-control order; the controls were lifted on June 30 and access was restored on July 1. That sequence is documented history — Simon Willison watched the Anthropic API start returning errors in real time, and Anthropic published its own statement about the suspension and restoration.

What is new is how Opus 5 handles questions about it. The model’s training cutoff is May 2026, so the shutdown and its reversal fall outside what it learned. The roughly 200,000-character Claude Opus 5 system prompt that red-teamer Pliny the Liberator posted on July 25 — a document Anthropic has not confirmed matches production — records the full timeline and instructs the model to answer questions about the suspension plainly rather than deny it, pointing users to Anthropic’s official statement when it cannot confirm the current status. It is a concrete example of a vendor writing post-cutoff history into the prompt so the model does not improvise about its own family.

When not to use it

Don't reach for Opus 5 when you specifically need an independently-audited SWE-bench Verified coding number to cite today — that figure isn't published yet, and Fable 5 (95.0%) still owns it. Don't assume every effort setting is #1; the leaderboard-topping 61 is the max setting, and lower settings trade some quality for cost, so benchmark the effort level you'll actually run. And for pure high-volume, latency-sensitive text with no reasoning need, a cheaper efficiency-tier model (for example Gemini 3.6 Flash at $1.50 / $7.50) will still be cheaper per token — Opus 5 is a reasoning and agent model, not a bulk-text one.

FAQ

How much does Claude Opus 5 cost?

$5 per 1M input tokens and $25 per 1M output tokens — the same as Opus 4.8, and half of Fable 5's $10 / $50. [Anthropic]

Is Opus 5 really the #1 model?

On the Artificial Analysis Intelligence Index, yes as of this writing: Claude Opus 5 (max effort) leads at 61 out of 170 models, ahead of Fable 5 (60). That is an independent score, not a vendor claim. Other leaderboards and your own workload may rank differently, and rankings change as models are added.

What is the effort dial?

A per-request setting (low/medium/high/max) that trades tokens — and therefore cost and latency — for capability. Max tops the leaderboard; cheaper settings still beat most rivals per Anthropic. API users set it as a parameter.

Is Opus 5 better than Fable 5?

On the independent Intelligence Index (61 vs 60) and Anthropic's Frontier-Bench v0.1 (43.3% vs 33.7%), yes — at half the price. Fable 5 still leads the independently-measured SWE-bench Verified (95.0%), so for pure coding-accuracy citations Fable 5 keeps that specific crown.

How does Opus 5 compare to Opus 4.8?

Same $5/$25 price, materially higher vendor benchmarks (Frontier-Bench v0.1 43.3% vs 18.7%) and gains across every tracked life-sciences eval, plus the new effort dial and safer alignment. For existing 4.8 users it's a free-of-price-increase upgrade.

How safe is it?

Anthropic calls it the most aligned Opus model and the least susceptible to misuse, and expects its cyber classifiers to intervene about 85% less often than Fable 5's — typically fewer false refusals on legitimate work.

Where can I use Claude Opus 5?

Claude.ai (default on Max, strongest on Pro), the Claude API, Claude Code, and Claude Cowork — live since July 24, 2026.

What's the context window?

Opus-family 1M-token class (Opus 4.8 is 1M with no long-context surcharge). Anthropic hasn't separately broken out an Opus 5 figure, so treat it as 1M-class until confirmed.

Does it support images or video?

Opus 5 is primarily a text-and-tools reasoning model. For heavy multimodal (image/video) understanding, a model like Gemini 3.1 Pro is purpose-built; use Opus 5 for text, code, and agents.

Bottom line

Claude Opus 5 turned out to be more than a pricing move. It is priced like a workhorse — $5/$25, half of Fable 5 and the same as Opus 4.8 — but it now leads the independent Artificial Analysis Intelligence Index at 61 (max effort), backed by strong vendor benchmarks (Frontier-Bench v0.1 43.3%, AutomationBench, OSWorld 2.0), a genuinely useful effort dial, and the safest alignment profile of any Opus. The one caveat is a still-missing independent SWE-bench Verified number, where Fable 5 (95.0%) keeps the coding-accuracy citation. For existing Claude users the upgrade is low-risk and doesn't raise the bill; for everyone else, Opus 5 is now the model to beat on quality-per-dollar — validate it on your own workload, and route it alongside every competitor through one OpenAI-compatible endpoint at OrcaRouter.