Claude Opus 5 Benchmarks: What the Numbers Actually Say

Anthropic launched Claude Opus 5 on July 24, 2026, with benchmark claims that read like a clean sweep: more than double Opus 4.8 on one eval, three times the next-best model on another, and past Fable 5 at a third of the cost on a third.

Try Apidog today

That does not make the claims wrong. It does mean you should read them like any vendor launch chart: carefully, with attention to what is measured and what is omitted. This guide puts every published Opus 5 benchmark claim in one attributed table, explains what each number can and cannot support, identifies the ceiling Anthropic places above its own model, and ends with an eval plan you can run in Apidog.

For the model itself (claude-opus-5, 1M context, 128k max output, May 2026 knowledge cutoff), start with what Claude Opus 5 is. The primary source below is Anthropic’s Claude Opus 5 launch post.

Every published claim, in one table

Here is the complete set of benchmark statements Anthropic made at launch. The comparison column matters as much as the result column because several claims are relative to another model instead of reporting a raw score.

Benchmark	Anthropic’s claim	Stated as	Compared against
Frontier-Bench v0.1	Surpasses all other models; more than doubles Opus 4.8’s score, at a lower cost per task	Ratio plus cost claim	Opus 4.8 and the field
ARC-AGI 3	Roughly 3x the next-best model’s score	Ratio	Unnamed next-best model
OSWorld 2.0	Surpasses Fable 5 at one-third the cost	Ordinal plus cost claim	Fable 5
CursorBench 3.2	Within 0.5% of Fable 5’s peak, at half the price	Gap plus cost claim	Fable 5
Zapier AutomationBench	Pass rate roughly 1.5x the next-best model	Ratio	Unnamed next-best model
Organic chemistry	+10.2 points over Opus 4.8	Absolute delta	Opus 4.8
Protein analysis	+7.7 points over Opus 4.8	Absolute delta	Opus 4.8
Cybersecurity exploitation	Behind Mythos 5	Ordinal	Mythos 5
Autonomous biology research	Behind Mythos 5	Ordinal	Mythos 5

Source for all of the above: Anthropic’s launch post and model documentation. No third party has published a reproduction of these results at the time of writing.

Two things stand out before you inspect individual numbers:

• Only two of the nine rows report an absolute point change.
• Anthropic includes the final two rows, where its new flagship loses.

The ratio problem, and why it matters

Four claims—Frontier-Bench, ARC-AGI 3, AutomationBench, and, to a degree, CursorBench—are expressed as ratios or gaps rather than scores. That is a meaningful limitation.

A ratio hides the denominator. “Roughly 3x the next-best model” on ARC-AGI 3 means very different things depending on the baseline:

• If the next-best score is 8%, then 3x is 24%.
• If the next-best score is 25%, then 3x is 75%.

Both can accurately be described as “3x,” but they imply radically different capability levels. Anthropic did not publish the underlying score.

Doubling is easier near the bottom of a scale. “More than doubles Opus 4.8’s score” on Frontier-Bench v0.1 has the same issue. If Opus 4.8 scored in the single digits, doubling may still be a modest absolute gain. If it scored 40%, doubling would be extraordinary.

The v0.1 label is relevant: this is a new benchmark with no published track record, community reproductions, or established saturation point.

“Next-best model” is unnamed. ARC-AGI 3 and AutomationBench do not specify the competitor. Whether that competitor is Fable 5, Mythos 5, or another model materially changes how you should interpret the ratio.

Gaps need units. “Within 0.5% of Fable 5’s peak” on CursorBench 3.2 does not clarify whether this means:

• 0.5 percentage points, or
• a 0.5% relative difference.

“Peak” also suggests a best-configuration run rather than default settings. On coding benchmarks, the difference between a default effort level and a maxed-out configuration can be significant. See this Fable 5 benchmarks breakdown for how that model’s results were framed.

The science claims are the clearest in the set because they use absolute deltas:

• +10.2 points on organic chemistry
• +7.7 points on protein analysis

You still do not get the starting score, but you know exactly how far the model moved.

The cost-per-task claims are configuration-dependent

Three claims combine capability and cost:

• Lower cost per task on Frontier-Bench
• One-third the cost on OSWorld 2.0
• Half the price on CursorBench 3.2

The list-price portion is verifiable. Opus 5 costs $5 per million input tokens and $25 per million output tokens—identical to Opus 4.8 and exactly half of Fable 5’s $10/$50 pricing.

That information comes from Anthropic’s pricing page. It is not itself a benchmark result. For cache writes, batch rates, and fast-mode pricing, see the Opus 5 pricing breakdown.

Cost per task is different from cost per token. It depends on:

• Total input and output tokens
• The selected effort level
• Whether thinking is enabled
• The number of turns in the agentic loop
• The number of subagents spawned
• Cache usage

Anthropic’s prompting guidance notes that Opus 5 produces longer default responses than Opus 4.8, delegates to subagents more readily, and expands task scope more often. Each behavior can increase token consumption in real workloads.

This does not make the cost claims false. Half the list price is still half the list price, and the Opus 5 vs Fable 5 comparison explains where that price difference can pay off.

It does mean that cost-per-task is a result of a specific configuration. Your configuration will differ, so measure it.

The ceiling Anthropic names itself

The most useful statement in the launch material is not a win: Anthropic says that Opus 5 still trails Mythos 5 on cybersecurity exploitation and autonomous biology research.

Fable 5 also retains the “most capable widely released model” designation.

That distinction matters. Opus 5 is not at the top of the Claude stack.

A more accurate summary is:

Frontier-class capability at half the frontier price, with a named ceiling above it.

For frontier security research or autonomous scientific work, Anthropic’s benchmark answer remains Mythos-class. See the Mythos-class model explainer and Fable 5 vs Mythos 5.

There is an operational implication, too. Anthropic shipped a fallbacks: "default" option, behind the server-side-fallback-2026-07-01 beta header, that falls back to Opus 4.8 when Opus 5 refuses a cyber-category request.

A model with tighter cyber refusals needs an escape hatch. The existence of that fallback tells you something a benchmark chart does not.

What the benchmarks do not cover

Benchmarks measure task completion. They do not measure the product behavior that determines whether a model works in your application.

• Behavioral drift during migration. Opus 5 verifies its own work without being prompted. If your prompts still include “double-check your answer” instructions from Opus 4.8, you may get over-verification and wasted tokens. Anthropic’s Opus 5 prompting guide recommends removing them.

• Output length. Default responses and written deliverables are longer than Opus 4.8. Lowering effort reduces thinking tokens, not necessarily visible output length. Ask for concise output explicitly.

• Failures with thinking disabled. With thinking: {type: "disabled"}, tool calls can occasionally appear as plain text instead of executing. Those text artifacts can poison later turns in an agentic loop. Internal XML tags can also leak into visible output.

• Effort recalibration. Effort levels were recalibrated for Opus 5. Anthropic recommends re-sweeping levels instead of carrying Opus 4.8 settings forward. See the effort parameter guide.

• The 400 error you will hit. Combining thinking: {type: "disabled"} with xhigh or max effort returns a per-request 400 error. This is a migration issue, not a capability issue. It is covered in the Opus 4.8 to Opus 5 migration guide.

What to actually test yourself

Treat vendor benchmarks as a starting hypothesis. Run this compact eval against your own workload.

1. Build a task set from production history

Collect 20 to 50 real tasks, such as:

• Closed support tickets
• Merged pull requests
• Bug reports
• Incident follow-ups
• Internal requests your model would have handled

Real inputs are better than synthetic prompts because they include your actual ambiguity, formatting, conventions, and edge cases.

2. Pin the configuration

Record the exact configuration for every run:

• Model ID: claude-opus-5
• output_config.effort
• Whether thinking is enabled
• max_tokens
• Tool definitions and versions
• System prompt version

Without a pinned configuration, later comparisons are not meaningful.

3. Score cost per resolved task

Do not optimize for cost per token alone. For every response, read and store the usage block:

• Input tokens
• Output tokens
• Cache-read tokens
• Cache-write tokens

Then calculate:

cost_per_resolved_task = total_spend / tasks_that_pass_acceptance

This is the metric Anthropic’s cost-per-task charts imply. It is also the metric that matters in production.

4. Run an effort sweep

Run the same task set at:

low
medium
high
xhigh

The recalibration means low and medium are stronger on Opus 5 than on earlier Opus models. A lower effort level may pass your acceptance criteria at a lower cost.

At higher effort levels, watch max_tokens closely.

5. Test the agentic loop, not only single turns

Many launch benchmarks are agentic: OSWorld, CursorBench, and AutomationBench.

A single-turn spot check will not reproduce these conditions. Run multi-turn tasks with your actual tools attached, and track:

• Successful completion
• Number of turns
• Tool-call success rate
• Retries
• Refusals
• Total token usage
• Time to completion

6. Compare with your incumbent model

Run your existing model through the same harness:

• Opus 4.8
• Sonnet 5
• Another provider’s model
• Your current production configuration

A relative result on your own data is more useful than an absolute score on someone else’s benchmark.

7. Log refusals separately

Cyber-adjacent tasks may trigger refusals more frequently than they did on Opus 4.8. Count refusals separately before deciding whether to implement the fallback header.

For a previous launch comparison methodology, see the Sonnet 5 benchmarks piece. The Opus 5 API guide includes request shapes.

Running the eval in Apidog

This eval is fundamentally API work: authenticated HTTP requests, JSON bodies, streaming responses, tool calls, and usage fields to capture on every call. That is what Apidog handles.

Use this setup:

1. Create a request for the Anthropic Messages endpoint.
2. Store the API key as an environment variable instead of embedding it in the request body.
3. Duplicate the request for each effort level so you can compare low through xhigh side by side.
4. Enable streaming when you need to inspect SSE events and see how thinking tokens accumulate before visible output.
5. Inspect tool-call payloads to verify that the model emits structured tool calls rather than describing calls in prose.
6. Add assertions for the usage fields so token totals and cache hits are captured automatically.
7. Save everything as a collection so the next model launch requires a model-ID swap rather than a rebuild.

Here is a base request:

curl https://api.anthropic.com/v1/messages \
  --header "x-api-key: $ANTHROPIC_API_KEY" \
  --header "anthropic-version: 2023-06-01" \
  --header "content-type: application/json" \
  --data '{
    "model": "claude-opus-5",
    "max_tokens": 8192,
    "output_config": {"effort": "medium"},
    "messages": [
      {"role": "user", "content": "Fix the failing test in this diff."}
    ]
  }'

Change one field per run, keep everything else fixed, and log the usage block every time. Download Apidog to follow this pattern with environments and assertions already wired up.

The honest summary

Opus 5’s benchmark story is strong: it is the largest jump over its predecessor Anthropic has claimed for an Opus release, at unchanged list pricing.

It is also entirely vendor-run, unreproduced as of July 25, 2026, and mostly expressed in ratios that hide their denominators.

Both statements can be true.

Treat the benchmark table as Anthropic’s hypothesis about its own model. Then get your own number. The gap between a launch chart and a production workload is where model migrations are actually decided. For the rest of the changes, see the Opus 5 pillar guide.

FAQ

Are the Claude Opus 5 benchmarks independently verified?

No. As of July 25, 2026, every published Opus 5 benchmark result comes from Anthropic. No third-party lab or independent evaluator has released a reproduction. Treat them as vendor-reported figures.

What is the biggest Opus 5 benchmark claim?

Frontier-Bench v0.1, where Anthropic says Opus 5 more than doubles Opus 4.8’s score at a lower cost per task. Anthropic did not publish the underlying scores, only the ratio, and Frontier-Bench v0.1 is a new benchmark without an established track record.

Is Claude Opus 5 the most capable Claude model?

No. Fable 5 retains the “most capable widely released” designation, and Anthropic says Opus 5 still trails Mythos 5 on cybersecurity exploitation and autonomous biology research. Opus 5’s pitch is frontier-class capability at half of Fable 5’s price, not the top of the stack.

How much better is Opus 5 than Opus 4.8 on science tasks?

Anthropic reports +10.2 points on organic chemistry and +7.7 points on protein analysis over Opus 4.8. These are the two claims stated as absolute deltas rather than ratios, making them easier to interpret, although baseline scores were not published.

Does Opus 5 beat Fable 5?

On Anthropic’s numbers, Opus 5 surpasses Fable 5 on OSWorld 2.0 at one-third the cost and lands within 0.5% of Fable 5’s peak on CursorBench 3.2 at half the price. Fable 5 still holds the overall capability designation. See the full comparison.

What should I benchmark myself?

Measure cost per resolved task across 20 to 50 real tasks from your backlog. Run multiple effort levels, attach your actual tools, enable multi-turn loops, and compare results against your current model using the identical harness.