Anthropic shipped Claude Opus 5 on July 24, 2026, and Claude Sonnet 5 on June 30, 2026. Both are current-generation models with a 1M-token context window, 128k max output, adaptive thinking enabled by default, and an effort parameter. The practical decision is not just capability—it is cost per completed task.
Sonnet 5 is currently priced at an introductory $2 per million input tokens and $10 per million output tokens. On September 1, 2026, it moves to $3 / $15. Opus 5 remains at $5 / $25. That makes Opus 2.5x more expensive today, but only about 1.67x more expensive after the Sonnet price change. For Q4 planning, use the September rates.
This guide compares specifications, cost per task, workload fit, and a practical test setup. You can run both models side by side in Apidog and compare the returned usage fields using your own prompts.
Pricing: account for the September change
Anthropic publishes the following base rates on its pricing page.
| Model | Input / MTok | Output / MTok | Status |
|---|---|---|---|
| Claude Opus 5 | $5.00 | $25.00 | Standard |
| Claude Sonnet 5, through Aug. 31, 2026 | $2.00 | $10.00 | Introductory |
| Claude Sonnet 5, from Sept. 1, 2026 | $3.00 | $15.00 | Standard |
Before building a cost model, include Opus 5's additional pricing options:
- Prompt caching: $6.25/MTok for a 5-minute cache write, $10/MTok for a 1-hour write, and $0.50/MTok for cache hits.
- Batch API: $2.50 input / $12.50 output per MTok, a flat 50% discount.
- Fast mode: $10 input / $50 output per MTok for 2.5x output speed. This is a research preview, first-party API only, and cannot be combined with batch.
-
US inference geography:
inference_geo: "us"applies a 1.1x multiplier to every token category.
See the Opus 5 pricing guide and Sonnet 5 pricing breakdown for worked examples.
One useful baseline: Opus 5 costs exactly what Opus 4.8 cost. Anthropic did not raise the Opus price for this generation.
Specs: almost the same platform surface
| Spec | Claude Opus 5 | Claude Sonnet 5 |
|---|---|---|
| Model ID | claude-opus-5 |
claude-sonnet-5 |
| Context window | 1M tokens | 1M tokens |
| Max output | 128k; 300k on Batch API with beta header | 128k |
| Adaptive thinking | On by default | On by default |
| Effort levels | Five, up to max; default high
|
low, medium, high, xhigh
|
| Knowledge cutoff | May 2026 | January 2026 |
| Priority Tier | Not supported | Not available |
| Prompt caching, tools, batch, vision | Supported | Supported |
| Tokenizer | 4.7-generation tokenizer | 4.7-generation tokenizer |
Three details affect implementation decisions:
Knowledge cutoff differs by four months. Opus 5 includes knowledge through May 2026; Sonnet 5 stops at January 2026. Use retrieval for changing library, API, and product information whenever possible. Without retrieval, Opus has the recency advantage.
Neither model supports Priority Tier. Teams migrating from Opus 4.8 should not treat either model as a latency-guarantee replacement. Review the Opus 4.8 to Opus 5 migration guide before upgrading.
Tokenization is identical between these models. Do not try to choose based on tokenizer efficiency. Both use the same tokenizer.
Measure cost per completed task, not cost per token
Anthropic states that Claude 4.7 and later models use a newer tokenizer that produces roughly 30% more tokens for the same text than previous models. That comparison is against Sonnet 4.6 and earlier—not between Opus 5 and Sonnet 5.
For the same prompt:
- Opus 5 and Sonnet 5 produce the same input token count.
- Their published token prices are directly comparable.
- If you are migrating from Sonnet 4.6 or earlier, account for the token-count increase.
The bigger cost variable is usually output volume. Anthropic's prompting guidance says Opus 5 can produce longer visible responses and written deliverables than prior Opus models. Lowering effort can reduce thinking, but does not reliably shorten visible output.
Because output tokens cost five times more than input tokens on both models, enforce concise output in the prompt:
Return only the implementation plan.
Use at most 8 bullet points.
Do not restate the problem.
Do not include background explanation unless required for a decision.
Use Anthropic's token counting endpoint to measure exact input tokens, then sample real outputs and retries.
For a daily workload of 500k input tokens and 40k output tokens:
| Scenario | Daily cost | Ratio vs. Sonnet 5 |
|---|---|---|
| Opus 5 | $3.50 | Baseline |
| Sonnet 5, introductory $2/$10 | $1.40 | Opus is 2.5x |
| Sonnet 5, standard $3/$15 | $2.10 | Opus is 1.67x |
| Opus 5 producing 30% more output | $3.80 | Opus is 1.81x |
Retries can reverse the apparent savings. At standard pricing:
- One Opus 5 run: $3.50
- One Sonnet 5 run: $2.10
- Two Sonnet 5 attempts: $4.20
If Sonnet needs a second attempt where Opus succeeds once, Sonnet costs more before accounting for human review. See how to cut your Claude API bill for general cost controls.
Where Opus 5 earns the premium
Anthropic's Opus 5 launch benchmarks are vendor-run results published by Anthropic. As of July 25, 2026, they have not been independently reproduced. Treat them as sourced vendor claims, not neutral measurements.
Anthropic reports that Opus 5:
- Surpasses other models on Frontier-Bench v0.1 and more than doubles Opus 4.8's score at lower cost per task.
- Scores roughly 3x the next-best model on ARC-AGI 3.
- Surpasses Fable 5 on OSWorld 2.0 at about one-third the cost.
- Lands within 0.5% of Fable 5's CursorBench 3.2 peak at half the price.
- Achieves a Zapier AutomationBench pass rate around 1.5x the next-best model.
- Improves organic chemistry by 10.2 points and protein analysis by 7.7 points over Opus 4.8.
Sonnet 5's launch figures were below Opus 4.8 on SWE-bench Pro, Terminal-Bench 2.1, and OSWorld-Verified. Its original positioning was near-Opus 4.8 agentic and tool-use performance at much lower cost. Opus 5 changes that comparison because the higher tier is now materially stronger than Opus 4.8.
See the Opus 5 benchmarks breakdown and Sonnet 5 benchmarks post for attribution and caveats.
Choose Opus 5 when one or more of these conditions apply:
- Long-horizon agentic coding: multi-file refactors, code migrations, and tool loops spanning dozens of actions.
- High cost of failure: production migrations, security-adjacent code, or code requiring expensive human review.
- Computer use: desktop and browser automation where multi-step reliability matters.
- Hard scientific or analytical reasoning: especially chemistry and protein-related workloads.
- Recency-sensitive tasks without retrieval: when May 2026 knowledge matters.
Where Sonnet 5 is enough
For most production traffic, Sonnet 5 is likely the default choice. Paying for quality that does not change your acceptance result is wasted spend.
Use Sonnet 5 for:
- High-volume, small requests: classification, extraction, routing, tagging, and summarization.
- Automatically validated outputs: schema checks, deterministic parsers, test suites, or other cheap failure detection.
- Latency-sensitive user flows: especially retrieval-augmented chat.
- Prototypes and internal tools: start cheaper, then upgrade only after measuring a meaningful quality difference.
Sonnet 5 is also the default model on the free Claude plan, so you can test prompts before implementing an integration. Opus 5 is available to Pro subscribers and is the default for Max. See the Opus 5 free access guide for the available paths.
Decision table
| Workload | Pick | Why |
|---|---|---|
| Multi-file refactor or migration in an agent loop | Opus 5 | Compounding errors make per-step reliability the dominant cost |
| High-volume classification or extraction | Sonnet 5 | Both hit the quality ceiling; price wins |
| Customer-facing chat with retrieval | Sonnet 5 | Retrieval closes the cutoff gap; latency matters |
| Computer use or desktop automation | Opus 5 | Anthropic's OSWorld 2.0 claim is strongest here |
| Overnight document processing | Sonnet 5, or Opus 5 Batch API | Batch pricing can make Opus close to Sonnet's standard rate |
| Security review or production-critical code | Opus 5 | Cost of a miss exceeds the token delta |
| Scientific or quantitative analysis | Opus 5 | Chemistry and protein-analysis gains are strongest |
| Prototyping and internal tools | Sonnet 5 | Measure before upgrading |
| Spring 2026 knowledge without retrieval | Opus 5 | May 2026 cutoff vs. January 2026 |
| Latency-sensitive interactive UX | Sonnet 5 | Or Opus 5 fast mode if depth is non-negotiable |
The Batch API deserves special attention. Opus 5 batch pricing is $2.50 input / $12.50 output per MTok. That is below Sonnet 5's September standard input rate and slightly above its output rate. If asynchronous processing works for your workload, batch pricing can make the tier choice much less important.
What neither model is
Opus 5 is not Anthropic's top model. Fable 5 remains Anthropic's most capable widely released model at $10 / $50.
Opus 5 also trails Mythos 5 on cybersecurity exploitation and autonomous biology research, according to Anthropic.
The practical framing is: Opus 5 provides frontier-class capability at half the frontier price, but a higher capability tier still exists. If your workload requires that tier, compare Opus 5 vs. Fable 5 instead.
Run the comparison in Apidog
Benchmarks are useful context, but your prompts, tools, and acceptance criteria determine the actual choice. Both models use the same Messages endpoint, so the test setup differs by one field: model.
Set up the test in Apidog:
- Create a request for the Anthropic Messages endpoint.
- Store
ANTHROPIC_API_KEYas an environment variable. Do not put the key directly in the request body. - Save two request variants with the same prompt:
"model": "claude-opus-5""model": "claude-sonnet-5"
- Capture the
usageobject from each response. - Add assertions for response shape, required fields, and expected output limits.
- Run the same task set against both variants.
- Compare pass rate, retries, output size, latency, and token usage—not just per-request cost.
- For Opus 5, repeat the test at several
output_config.effortlevels.
Example 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": 4096,
"messages": [
{
"role": "user",
"content": "Refactor this handler to use async I/O."
}
]
}'
Use assertions to catch failures early. For example, validate that the response is complete, has expected content, and does not exceed a cost or token threshold:
// Example test logic
const body = pm.response.json();
pm.test("Response contains content", () => {
pm.expect(body.content).to.be.an("array").that.is.not.empty;
});
pm.test("Response was not truncated", () => {
pm.expect(body.stop_reason).to.not.eql("max_tokens");
});
pm.test("Usage was returned", () => {
pm.expect(body.usage).to.have.property("input_tokens");
pm.expect(body.usage).to.have.property("output_tokens");
});
Two implementation gotchas:
- Adaptive thinking is enabled by default on both models, and
max_tokenscaps thinking plus visible response together. A limit that worked for an older model can now truncate output. - On Opus 5, combining
thinking: {type: "disabled"}with effortxhighormaxreturns a 400 response.
The effort parameter guide explains how to run an effort sweep. You can also download Apidog to execute both request variants side by side, and use the Opus 5 API walkthrough for the full request flow.
FAQ
Is Claude Opus 5 worth 2.5x the price of Sonnet 5?
For long-horizon agentic coding, computer use, and work that is expensive to get wrong, often yes. For high-volume classification, extraction, and chat, usually not. The multiple also changes after September 1, 2026, when Sonnet 5 moves to $3 / $15 and Opus becomes about 1.67x the cost.
When does Claude Sonnet 5 pricing change?
The introductory $2 / $10 rate runs through August 31, 2026. On September 1, 2026, Sonnet 5 moves to $3 / $15. Opus 5 remains $5 / $25.
Do Opus 5 and Sonnet 5 tokenize text differently?
No. Both use the 4.7-generation tokenizer, so the same text creates the same input token count. The roughly 30% token increase applies when comparing Claude 4.7+ models to Sonnet 4.6 and earlier models.
Can I switch between Opus 5 and Sonnet 5 without code changes?
Mostly. Both use adaptive thinking by default and reject non-default sampling parameters. Watch for effort-level differences: Opus 5 includes max and defaults to high. Also, Opus 5 returns a 400 when thinking is disabled with xhigh or max effort.
Which model does Anthropic recommend by default?
Anthropic's documentation recommends starting with Opus 5 if you are unsure. Treat that as a capability recommendation, not a cost recommendation. Measure whether Sonnet 5 produces the same accepted result for less. The Sonnet 5 vs. Opus 4.8 comparison remains useful background.
For full specifications, benchmark attribution, and availability details, start with what Claude Opus 5 is, then review Anthropic's Opus 5 launch post and models overview.
Top comments (0)