How to Use Claude Sonnet 5: Effort, Cost, and Limits

Anthropic released Claude Sonnet 5 on June 30, 2026 with an unusually direct pitch: near-Opus intelligence on coding and agentic work, at Sonnet pricing^[1]. The official benchmarks back up most of it.

Knowing how to use Claude Sonnet 5 well is mostly about two things the launch post does not lead with. The effort ladder is the lever that makes the cost claim real, and a denser tokenizer means the same prompt now costs about 30% more tokens than it did on Sonnet 4.6, so a migration that only swaps the model string will quietly change your bill and can truncate your output^[1].

This guide covers the benchmarks, where Sonnet 5 ties and where it trails Opus 4.8, the four API changes that break carried-over code, the pricing including the September increase, and who should actually run it.

TL;DR

• Near-Opus on knowledge work. GDPval-AA v2 Elo of 1618 against Opus 4.8's 1615, a statistical tie^[1].
• Still trails on the hard tail. SWE-bench Pro 63.2% against 69.2%, and USAMO 2026 79.5% against 96.7%^[1].
• Introductory pricing runs through August 31, 2026: $2 input / $10 output per million tokens, rising to $3 / $15 on September 1^[1].
• First Sonnet with the full effort ladder: low, medium, high (default), xhigh, max^[1].
• The tokenizer is about 30% denser. Per-token prices are unchanged, so an identical request costs more than on Sonnet 4.6. Re-baseline before you migrate^[1].
• budget_tokens and sampling parameters now return 400^[1].

What Claude Sonnet 5 is

Claude Sonnet 5 (claude-sonnet-5) succeeds Sonnet 4.6 as a drop-in replacement. Anthropic calls it the most agentic Sonnet model yet, built to make plans, drive browsers and terminals, and run autonomously for long stretches at a level that recently required larger and more expensive models^[1].

It shipped as the default model for Free and Pro users on claude.ai, and simultaneously on the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry in preview. It carries a 1M-token context window by default and 128K maximum output^[1].

Benchmarks

Two things stand out. Against Sonnet 4.6 this is a clean sweep with real jumps: Terminal-Bench climbs 13.4 points and the knowledge-work Elo rises 223. And on knowledge work Sonnet 5 nominally edges Opus 4.8, 1618 against 1615, which is a tie, while landing within roughly two to six points of the flagship on agentic coding and computer use^[1].

That "within a few points, at a third of the cost" gap is the entire value proposition.

Deeper numbers from the system card: 85.2% on the classic 500-problem SWE-bench Verified subset, 78.3% on SWE-bench Multilingual across nine languages, and 38.8 on Cognition's FrontierCode v1, more than double Sonnet 4.6's 15.1. Cursor's independent measurement puts it at 61.2% on CursorBench against Sonnet 4.6's 49% and Opus 4.8's 63.8%. BrowseComp reaches 84.7% single-agent and 86.6% multi-agent^[1].

One honest soft spot: on USAMO 2026 Sonnet 5 scores 79.5%, a big jump over Sonnet 4.6's 55.0% but far behind Opus 4.8's 96.7%^[1]. Heavy formal-proof mathematics is not this model's job.

The effort ladder is the cost lever

Sonnet 5 is the first Sonnet-tier model to expose the full ladder: low, medium, high (the default), xhigh, and max^[1]. This is not a cosmetic knob. It is what makes the Opus-quality-at-Sonnet-cost claim real.

The BrowseComp curve is the clearest illustration. Sonnet 5's accuracy climbs from roughly 60% at low to about 85% at max, and it reaches Opus-4.8-class accuracy at a markedly lower cost per task. Opus 4.8 still tops out slightly higher, but Sonnet 5 gets most of the way there for far less money, which is exactly what a high-volume agent loop needs^[1].

Computer use tells a more sober story.

On OSWorld-Verified, Opus 4.8 stays ahead at every effort level. Sonnet 5 still beats Sonnet 4.6 comfortably and remains far cheaper, but if pixel-accurate GUI control is your core workload the flagship keeps a genuine edge^[1].

The practical rule: xhigh for the hardest long-horizon coding and agentic tasks, high for most work, medium or low for latency-sensitive or simple jobs. Tune it per route rather than defaulting to max everywhere.

Against the cost-comparable tier

The system card publishes competitor numbers against GPT-5.5 and Gemini 3.5 Flash, which is the cost-comparable tier rather than the absolute frontier. Read it as "beats the models at its price point," not "beats everything"^[1].

Sonnet 5 wins most head-to-heads against GPT-5.5 and loses the one that gets quoted most: Terminal-Bench 2.1, where GPT-5.5 driving the Codex CLI scores 83.4 against 80.4. Against Gemini 3.5 Flash the sweep is broader, with AutomationBench the single exception. Note that AutomationBench scores are low across the board; these are unsaturated benchmarks, not solved ones^[1].

Where it ties Opus 4.8 and where it trails

Ties or edges ahead: GDPval-AA v2 knowledge work (1618 vs 1615), agentic search on BrowseComp at comparable accuracy for a given task cost, Real-World Finance v2 (1219 vs 1222), and the AA-Briefcase professional-work Elo (1393 vs 1352)^[1].

Trails: SWE-bench Pro (63.2 vs 69.2), Terminal-Bench 2.1 (80.4 vs 82.7), OSWorld (81.2 vs 83.4), HLE no-tools (43.2 vs 49.8), CursorBench (61.2 vs 63.8), Toolathlon (54.3 vs 59.9), and USAMO 2026 (79.5 vs 96.7)^[1].

The pattern is consistent. On open-ended knowledge work and agentic search, Sonnet 5 is effectively Opus-class. On the hardest coding, computer use, and formal math, Opus 4.8 keeps a real but modest lead. For workloads running many agent turns where cost dominates, Sonnet 5 is the rational default and Opus is the escalation path for the hard tail.

How to use Claude Sonnet 5: four changes that break old code

Sonnet 5 adopts the Opus 4.7/4.8 request surface, so migrating is more than a string swap^[1].

Adaptive thinking is on by default. Sonnet 4.6 ran without thinking unless asked. Sonnet 5 thinks adaptively out of the box, deciding per request how much reasoning a task warrants. You can still disable it explicitly when you need raw speed.

budget_tokens is removed. Manual thinking budgets now return a 400. Control depth through effort levels instead.

Sampling parameters are rejected. Non-default temperature, top_p, and top_k return a 400. Steer with prompting.

The tokenizer is about 30% denser. Sonnet 5 shares the tokenizer introduced with Opus 4.7, which turns the same text into roughly 30% more tokens. Per-token pricing is unchanged, so an equivalent request can cost more than it did on Sonnet 4.6, and a max_tokens limit tuned for the old tokenizer can now truncate output mid-thought.

That last one is the expensive surprise. Re-measure real spend on your own prompts rather than assuming the sticker price maps cleanly from Sonnet 4.6.

One capability addition worth noting: Sonnet 5 is the first Sonnet-tier model with a high-resolution image tier, up to 2576 pixels on the long edge and 4784 visual tokens, against the old 1568-pixel cap. It activates automatically, which matters for dense document, chart, and screenshot understanding^[1].

Pricing

Against Opus 4.8 at $5 / $25, Sonnet 5 is roughly a third of the cost at introductory rates and about 60% at standard rates^[1]. Layer in the effort dial and prompt caching and the gap widens for a well-tuned agent workload.

Two caveats. The denser tokenizer means you should recompute real spend rather than trusting the sticker comparison. And Priority Tier is not available on Sonnet 5^[1].

Safety behavior your code has to handle

Sonnet 5 is the first Sonnet-tier model shipping with real-time cybersecurity safeguards. Requests touching prohibited or high-risk cyber and biology topics can be refused, returned as a successful HTTP 200 with a refusal stop reason rather than an error^[1].

If you are building security tooling or life-sciences workflows where benign-but-adjacent requests might trip a classifier, handle that stop reason explicitly instead of assuming every 200 carries usable content. Anthropic also reports lower hallucination and sycophancy rates than Sonnet 4.6, which matters wherever output is trusted downstream without a human check^[1].

Who should use Claude Sonnet 5

It is the rational default for high-volume agentic workloads where cost per turn compounds: coding agents, browser and terminal automation, retrieval and research loops. It fits product builders embedding an LLM into a shipping feature, and long-context document work where the 1M window and low per-token price make bulk processing viable.

Reach for Opus 4.8 or Opus 5 instead when the task is in the hard tail: the most demanding coding, pixel-accurate computer use, or formal mathematics, where the flagship's few-point lead is worth the premium.

Running a Sonnet-tier model on reAPI

To be straightforward about coverage: Claude Sonnet 5 is not on the reAPI gateway today. The Sonnet-tier model available is Claude Sonnet 4.6, and the Claude models on the gateway are Opus 5, Opus 4.8, Opus 4.7, Sonnet 4.6, and Fable 5.

That matters less than it sounds like, because of where the prices actually land. Sonnet 5 lists at $3 / $15 from September. On reAPI, Claude Opus 5 runs $2.40 / $12.00 and Sonnet 4.6 runs the same. So the tier-above model costs less on the gateway than Sonnet 5's standard list price, and Opus 5 is the model Anthropic describes as delivering frontier intelligence at half the cost of Fable 5.

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_REAPI_KEY",
    base_url="https://api.reapi.ai/v1",
)

resp = client.chat.completions.create(
    model="claude-opus-5",
    messages=[{"role": "user", "content": "Plan and run this multi-step refactor."}],
    max_tokens=16000,
    stream=True,
)

The endpoint is OpenAI-compatible and the native Anthropic /v1/messages surface is available too. Rates and request shapes are on reapi.ai/models, and the Opus 5 reference is at reapi.ai/docs/claude-opus-5.

If your evaluation specifically requires Sonnet 5's behavior rather than its price point, call it through Anthropic directly.

FAQ

Is Claude Sonnet 5 better than Opus 4.8?

On open-ended knowledge work and agentic search they are effectively tied, and Sonnet 5 nudges ahead on GDPval-AA v2 (1618 vs 1615). On the hardest coding, computer use, and formal math, Opus 4.8 keeps a modest lead^[1].

How much does Claude Sonnet 5 cost?

$2 per million input tokens and $10 per million output through August 31, 2026, rising to $3 / $15 on September 1. Prompt caching cuts cached-context reads by roughly 90%, and the Batch API halves the rate^[1].

What is the xhigh effort level?

An effort setting between high and max, tuned for long-horizon agentic and coding tasks. Sonnet 5 is the first Sonnet-tier model to support the full ladder^[1].

Why did my costs go up after migrating from Sonnet 4.6?

The tokenizer is about 30% denser, so the same text becomes roughly 30% more tokens at unchanged per-token prices. Recount your prompts and raise max_tokens so long outputs are not truncated^[1].

Does Claude Sonnet 5 support a 1M-token context window?

Yes, 1M by default at the standard per-token rate with no long-context surcharge, and up to 128K output tokens^[1].

What breaks when migrating from Sonnet 4.6?

budget_tokens returns a 400, non-default sampling parameters return a 400, adaptive thinking is now the default, and the denser tokenizer changes both cost and truncation behavior^[1].

Can Claude Sonnet 5 control a computer?

Yes. It supports tool use, web search, web fetch, code execution, and computer use, scoring 81.2% on OSWorld-Verified against Opus 4.8's 83.4%^[1].

Is Claude Sonnet 5 available on reAPI?

Not currently. The gateway carries Claude Opus 5, Opus 4.8, Opus 4.7, Sonnet 4.6, and Fable 5. Opus 5 on reAPI costs less per token than Sonnet 5's standard list rate.

Choosing between a tier and a price point

Claude Sonnet 5 delivers on its central promise. For coding and agentic work it lands close enough to Opus 4.8 that, at a third of the cost during the introductory window, it becomes the sensible default and reserves the flagship for the hard tail. It wins most head-to-heads against GPT-5.5 and Gemini 3.5 Flash, with honest exceptions on Terminal-Bench, AutomationBench, and formal math.

Two things decide whether that arithmetic survives contact with your workload. The September price step takes it from a third of Opus to about 60%, and the denser tokenizer means the sticker comparison understates real spend. Measure both on your own prompts before committing. That is how to use Claude Sonnet 5 without discovering the tokenizer in your invoice, and it is also why the tier label matters less than the number you actually pay per completed task.

References

1. Anthropic. Introducing Claude Sonnet 5. Retrieved July 2026 from anthropic.com/news/claude-sonnet-5
2. Anthropic. Models overview — specs and pricing for all current Claude models. Retrieved July 2026 from platform.claude.com/docs/en/about-claude/models/overview
3. Anthropic. Pricing — token, cache, and batch rates. Retrieved July 2026 from platform.claude.com/docs/en/about-claude/pricing

Further reading

• reAPI. How to use Claude Opus 5. reapi.ai/blog/how-to-use-claude-opus-5
• reAPI. How to use Claude Opus 4.8. reapi.ai/blog/how-to-use-claude-opus-4-8
• reAPI. Model catalog. reapi.ai/models