by Anthropic · Claude 5 family · best for near-Opus agentic coding and computer use at $2/$10
Claude Sonnet 5 is Anthropic's mid-tier model, released June 30, 2026 as the direct successor to Sonnet 4.6 and built, in Anthropic's words, to be "the most agentic Sonnet model yet." On Anthropic's own evaluations it lands close to Opus 4.8 on agentic coding and computer use (SWE-bench Verified 85.2%, Terminal-Bench 2.1 80.4%, OSWorld-Verified 81.2%) at $2/$10 per million tokens, a price that was introductory at launch and became the permanent standard on August 10, 2026. It is the default model on the claude.ai Free and Pro plans. For a buyer: it is the Anthropic model to put on high-volume agent and coding traffic, but independent testing puts OpenAI's and Google's same-price workhorses ahead on general intelligence per dollar, so run it on your own tasks before you standardize.
budget_tokens) now returns a 400 error, and so do non-default temperature, top_p or top_k values.computer_toolset_20260801 computer use tool on the Claude API and Google Cloud, which Sonnet 4.6 does not.| Benchmark | Score | Source |
|---|---|---|
| Humanity's Last Exam | 43.2% | anthropic.com 2026-06-30T00:00:00.000Z |
| LMArena Elo | 1461 | lmarena.ai 2026-09-13T00:00:00.000Z |
| GPQA Diamond | 91.1% | artificialanalysis.ai 2026-09-21T00:00:00.000Z |
| Terminal-Bench | 80.4% | anthropic.com 2026-06-30T00:00:00.000Z |
| LMArena Coding Elo | 1520 | lmarena.ai 2026-09-13T00:00:00.000Z |
| SWE-bench Verified | 85.2% | anthropic.com 2026-06-30T00:00:00.000Z |
| Artificial Analysis Index | 38 | artificialanalysis.ai 2026-09-21T00:00:00.000Z |
Six personas, six verdicts — the same panel that reviews every product on TopReviewed.
“Sonnet 5 is the safe Anthropic default for agent traffic: close to Opus 4.8 on coding and computer use at $2/$10, on every major cloud.”
For an organization already on Claude, Sonnet 5 is an easy call. It replaces Sonnet 4.6 as a drop-in, costs less per token, ships on the Claude API, Bedrock, Google Cloud, Foundry and Claude Platform on AWS, and carries a retirement commitment to at least June 30, 2027. The strategic caveat is the market: GPT-5.6 Terra and Gemini 3.8 Flash score higher on independent general-intelligence tests at the same or lower price. Opus 5 is now Anthropic's recommended default for the hardest work. Standardize on Sonnet 5 for agentic volume, and route the frontier-hard decile to Opus 5.
“Anthropic's wedge is agentic reliability, not raw IQ: Sonnet 5 wins on computer use and low hallucination while rivals win the index charts.”
Sonnet 5 defends Anthropic's hold on agentic coding and computer use, the categories where enterprise spend is concentrating. It matches Opus 4.8 on GDPval-AA, beats Sonnet 4.6 across the board, and brings the stable computer use and browser toolsets to the mid tier. Its differentiation is reliability: Artificial Analysis finds it hallucinates far less than GPT-5.6 Terra or Sol. The positioning gap is general intelligence per dollar, where Gemini 3.8 Flash at $0.75/$3.75 and Grok 4.7 at $2/$6 now score higher. Timing is mixed: it shipped before GPT-5.6 reached GA and before Gemini 3.8, and now looks mid-pack on price.
“The $2/$10 rate card looks like a 33% cut, but the new tokenizer and heavy thinking at high effort make real savings closer to 10%.”
List price is attractive: $2/$10, cache reads at $0.20, batch at $1/$5, and a 1M window with no long-context premium. Two effects erode it. The tokenizer produces about 30% more tokens for the same text, so an equivalent request costs roughly 10-13% less than on Sonnet 4.6, not 33% less. And effort drives spend: running the Artificial Analysis Index cost $2,086 at high effort and $6,998 at max, against $2,501 for GPT-5.6 Terra at max with a higher score. Set effort per workload, cache aggressively, and push offline jobs to batch.
“A true drop-in: swap the model ID, strip temperature and budget_tokens, recount your tokens, and you get a much more persistent agent.”
Migration is one line plus three fixes: remove non-default sampling parameters, replace budget_tokens with adaptive thinking and an effort level, and recount prompts and max_tokens under the new tokenizer. In return you get the stable computer use and browser toolsets and an agent that testers say writes its own reproducing tests and verifies fixes unprompted. The Terminal-Bench gain was reproduced independently. Watch two gotchas: thinking text is omitted by default (set display: "summarized" to see it), and cyber safeguards can return stop_reason: "refusal" on security tooling.
“It finishes what it starts and makes things up less often, but default thinking adds seconds, so drop the effort level for quick chat.”
As the default on claude.ai Free and Pro, Sonnet 5 is most people's everyday Claude. It is more persistent than Sonnet 4.6, completing multi-step tasks that used to stall, and it is less sycophantic, with a lower hallucination rate on knowledge questions. The trade-offs are felt in chat. Adaptive thinking means roughly 12 seconds to the first word at the default high effort in API testing, and Anthropic reports no gain in creative writing or warmth, plus a more dismissive "wet blanket" streak. LMArena voters rank it slightly below Sonnet 4.6 for general chat.
“Near-Opus is Anthropic's framing, run at max effort on its own harnesses; independent indexes put Sonnet 5 behind GPT-5.6 Terra and Gemini 3.8 Flash.”
The headline numbers are vendor-run at max effort, averaged over five trials, on internal harnesses. OSWorld's harness also changed, re-rating Sonnet 4.6 to 78.5% and shrinking the apparent gain to 2.7 points. Independent data is thinner and less flattering: Artificial Analysis scores it 38, below Terra (42) and Gemini 3.8 Flash (41), at a far higher cost per run, and LMArena text voters put it below Sonnet 4.6. "Close to Opus 4.8" also fails on hard math (USAMO 79.5% vs 96.7%). In its favor, Terminal-Bench reproduced almost exactly, and its hallucination rate is much lower than OpenAI's.
budget_tokens return 400 errors, assistant prefill stays unsupported, and Priority Tier is unavailable.The full research notes behind this review — verified against primary sources.
Anthropic discloses no parameter count, layer count, attention design, expert count or training-compute figure, so those fields are null. The transparency hub says only that Sonnet 5 was pretrained on a proprietary mix of public web data, public and private datasets and synthetic data, then heavily post-trained, using AWS and Google Cloud compute with PyTorch, JAX and Triton. What is disclosed matters operationally: a 1M-token context window that is both the default and the only size, a 128K synchronous output cap, adaptive thinking steered by an effort parameter (low, medium, high, xhigh, max; default high), and a new tokenizer. Because each token covers less text, 1M tokens holds roughly 555k words on this tokenizer versus about 750k on Sonnet 4.6's.
Coding (9.4): SWE-bench Verified 85.2%, SWE-bench Pro 63.2%, SWE-bench Multilingual 78.3% and Terminal-Bench 2.1 80.4% (all vendor-reported); Artificial Analysis independently measured 80.5% on Terminal-Bench 2.1, and Cursor's own CursorBench put it at 61.2% against Opus 4.8's 63.8%. Human raters are less impressed: the LMArena coding score (1520) is statistically level with Sonnet 4.6's. Agentic (9.4): OSWorld-Verified 81.2%, BrowseComp 84.7% (single agent), Toolathlon 54.3%, GDPval-AA v2 1618, all vendor-run at max effort. Reasoning (8.9): GPQA Diamond 91.1% and HLE 41.3% in Artificial Analysis testing, up from Sonnet 4.6's 87.5% and 33.6% but below Opus 4.8. Math (8.7): USAMO 2026 79.5% (Sonnet 4.6 55.0%, Opus 4.8 96.7%) and ArXivMath 65.7% without tools. Long context (9.1): 1M tokens by default and 82% on Artificial Analysis's long-context reasoning test (AA-LCR), level with the best in its price band. Vision (8.7) and document work (8.5): MMMU-Pro 77.3% (Artificial Analysis), CharXiv Reasoning 77.0% without tools and 88.3% with them, GDP.pdf 67.5%/81.6%. Multilingual (9.0): Anthropic's published table, measured on Sonnet 4.5, keeps major languages within about 3.5% of English; the Sonnet 5 system card reports GMMLU, MILU and INCLUDE results as charts only. Instruction-following (9.0) and function-calling (9.2): early-access testers cite it staying on plan and following conventions, and it is a drop-in for Sonnet 4.6 tool definitions. Creative writing (8.2): Anthropic reports no improvement in creative mastery or warmth. Safety calibration (9.0): lower misaligned-behavior, hallucination and sycophancy rates than Sonnet 4.6. Realtime data (7.0): January 2026 cutoff, closed by the web search and web fetch tools.
| Benchmark | Score | vs Predecessor | vs Top Competitor | Source |
|---|---|---|---|---|
| SWE-bench Verified | 85.2% (vendor) | +5.6 vs Sonnet 4.6 (79.6% at its launch) | not on the public SWE-bench leaderboard yet | Anthropic system card |
| SWE-bench Pro | 63.2% (vendor) | +5.1 vs Sonnet 4.6 (58.1%) | behind Opus 4.8 (69.2%) | Anthropic |
| Terminal-Bench 2.1 | 80.4% (vendor); 80.5% (Artificial Analysis) | +13.4 vs Sonnet 4.6 (67.0%) | behind GPT-5.6 Terra max (88.0%) and Gemini 3.8 Flash (87.6%) in Artificial Analysis testing | Anthropic |
| OSWorld-Verified | 81.2% (vendor) | +2.7 vs Sonnet 4.6 (78.5%, re-run on the new harness) | behind Opus 4.8 (83.4%) | Anthropic |
| BrowseComp | 84.7% single agent, 86.6% multi-agent (vendor) | +8.5 vs Sonnet 4.6 (76.2%) | level with GPT-5.5 (84.4%) on Anthropic's table | Anthropic system card |
| Humanity's Last Exam | 43.2% no tools, 57.4% with tools (vendor) | +8.6 / +10.6 vs Sonnet 4.6 | Artificial Analysis: 41.3% vs Gemini 3.8 Flash 47.8% | Anthropic |
| GPQA Diamond | 91.1% | Sonnet 4.6 max 87.5% | behind Gemini 3.8 Flash (95.3%) and GPT-5.6 Terra max (92.5%) | Artificial Analysis |
| GDPval-AA v2 | Elo 1618 (vendor, as of 2026-06-17) | +223 vs Sonnet 4.6 (1395) | level with Opus 4.8 (1615) | Anthropic |
| USAMO 2026 | 79.5% (vendor) | +24.5 vs Sonnet 4.6 (55.0%) | well behind Opus 4.8 (96.7%) | Anthropic system card |
| MMMU-Pro | 77.3% | Sonnet 4.6 max 73.3% | behind Gemini 3.8 Flash (85.6%) | Artificial Analysis |
| AA-LCR (long context) | 82% | Sonnet 4.6 max 80% | level with GPT-5.6 Terra max (83%) and Gemini 3.8 Flash (81%) | Artificial Analysis |
| LMArena Text Elo | 1461 (+/-5, 35,301 votes) | below Sonnet 4.6 (1473) | behind Gemini 3.8 Flash (1493, preliminary) and GPT-5.6 Terra (1466) | LMArena |
| LMArena Coding Elo | 1520 (+/-7) | level with Sonnet 4.6 (1528, +/-6) | level with GPT-5.6 Terra (1518) | LMArena |
| Artificial Analysis Intelligence Index v4.3 | 38 (max effort; 32 at the default high effort) | +8 vs Sonnet 4.6 max (30) | behind GPT-5.6 Terra max (42), Gemini 3.8 Flash (41), Grok 4.7 (46) | Artificial Analysis |
(Anthropic's figures use adaptive thinking at max effort averaged over five trials, with internal harnesses; Terminal-Bench 2.1 runs on mini-SWE-agent rather than Terminus-2. The Artificial Analysis Index v4.3 is on a new scale and is not comparable with the v3-era "51" on the Sonnet 4.6 page. Anthropic's GDPval-AA v2 figure differs from Artificial Analysis's current GDPval-AA v2.1 run (1449). MMLU-Pro, AIME 2025, LiveCodebench, Aider Polyglot, MMMU, IFEval, tau-bench, MRCR and SimpleQA have no published Sonnet 5 figure and are null.)
Anthropic rates Sonnet 5 "Fast", its comparative label between Haiku 4.5 (Fastest) and Opus 5 (Moderate). Artificial Analysis measures generation speed at about 71 tokens/sec at the default high effort and 77 tokens/sec at max effort, above the 72.5 tokens/sec median for its price band. Time to first answer token is dominated by adaptive thinking, which is on by default: about 1.8 seconds at low effort, about 11.7 seconds at high effort, and 142.9 seconds at max effort, where the median for comparable reasoning models is 3.7 seconds. For chat and IDE loops, set effort to low or medium, or pass thinking: {type: "disabled"}. Reserve high and max for background agent work. For throughput, Gemini 3.8 Flash generates several times faster in the same testing.
| Surface | Cost | Notes |
|---|---|---|
| API input | $2 / 1M tok | Standard rate since August 10, 2026 (launch intro price made permanent) |
| API output | $10 / 1M tok | Thinking tokens bill as output |
| Cached input (read) | $0.20 / 1M tok | 0.1x base input |
| Cache write (5m / 1h) | $2.50 / $4 per 1M tok | 1.25x / 2x base input |
| Batch (in/out) | $1 / $5 per 1M tok | 50% off; up to 300K output with the beta header |
| US-only inference | 1.1x all token prices | inference_geo: "us"; Bedrock and Google Cloud regional endpoints carry a 10% premium |
| Web search tool | $10 / 1,000 searches | plus token costs |
| Direct UI | Free $0 · Pro $20/mo ($17/mo billed annually) · Max from $100/mo | Default model on Free and Pro |
| Free tier | claude.ai Free plan | usage limits on a rolling five-hour window |
| Rate limits | Start / Build / Scale tiers | Priority Tier not available on Sonnet 5 |
Tokenizer caveat: because the same text produces roughly 30% more tokens than on Sonnet 4.6, an equivalent request costs about 10-13% less, not the 33% the rate card suggests.
Proprietary: no open weights and no self-hosting. Available on the Claude API, Claude Platform on AWS, Amazon Bedrock (anthropic.claude-sonnet-5, including the InvokeModel API; the legacy Bedrock integration does not carry it), Google Cloud (Vertex AI) and Microsoft Foundry; the model ID is claude-sonnet-5 everywhere except Bedrock. Bedrock offers global and regional endpoints, and Google Cloud offers global, multi-region and regional endpoints, with the 10% regional premium. First-party US-only routing is available at 1.1x. Zero data retention is supported for organizations with ZDR agreements.
Deployed under Anthropic's Responsible Scaling Policy with CB-1 protections: real-time classifier guards for chemical and biological misuse that Anthropic describes as equal to or stronger than its historical ASL-3 protections. It is also the first Sonnet with real-time cyber safeguards on by default (the same, less strict set used on Opus 4.7 and 4.8); blocked requests return HTTP 200 with stop_reason: "refusal", and security teams can apply to the Cyber Verification Program. In Anthropic's testing it never produced a working Firefox exploit, and it shows lower rates of misaligned behavior, hallucination and sycophancy than Sonnet 4.6, although its misaligned-behavior rate is still higher than Opus 4.8's. Prompt-injection resistance is weaker than Opus 5: a 2.25% attack success rate in computer-use scenarios and 5.9% within 15 attempts on Anthropic's indirect-injection benchmark. API data is not used for training without express permission. HIPAA-ready access with a BAA, SOC 2 Type I and II, ISO 27001:2022 and ISO/IEC 42001:2023 apply.
Official SDKs in Python, TypeScript, Java, Go, Ruby, C# and PHP, with first-class support in Claude Code, the Claude Agent SDK and Claude Managed Agents (including Dreams). Model-agnostic frameworks such as LangChain, LlamaIndex and the Vercel AI SDK reach it through the standard Anthropic provider. It is the default model on claude.ai Free and Pro and runs in Claude Cowork. Named launch customers include Lovable, Kiro, ClickHouse, Pace and Eve, and Cursor benchmarked it in its production harness. LMArena shows 35,301 text votes by September 13, 2026, a mainstream adoption signal.
Yes for new builds and most agent workloads: it scores higher on every Anthropic benchmark and costs less per token. Before switching, remove non-default sampling parameters and budget_tokens, and recount tokens under the new tokenizer.
Yes. It launched as introductory pricing, and on August 10, 2026 Anthropic made it the standard price and cancelled the planned September 1 increase to $3/$15.
Sonnet 5 uses the newer tokenizer, which produces about 30% more tokens for the same text (1.0-1.35x by content). Budgets, max_tokens limits and cost models from Sonnet 4.6 need to be re-measured.
Use Opus 5 for the hardest reasoning, math and long-horizon work. In Artificial Analysis testing, Opus 5 at medium effort (44.8) outscored Sonnet 5 at max effort (38) for less total cost, so max-effort Sonnet is rarely the efficient choice.
Adaptive thinking is on by default at high effort. Use effort: "low" or "medium", or disable thinking, for interactive paths.
Possibly. Real-time cyber safeguards are on by default and return stop_reason: "refusal" on high-risk requests. Legitimate security teams can join the Cyber Verification Program, and Anthropic recommends Opus 4.8 where reduced guardrails are needed.
Yes. API data is not used for training without express permission, ZDR and HIPAA-ready access (BAA) are available, Anthropic holds SOC 2 Type I and II, ISO 27001 and ISO 42001, and US-only inference is available at 1.1x.
The price twin at $2/$12. Terra leads on the Artificial Analysis Index (42 vs 38 at max effort), GPQA and Terminal-Bench 2.1 in that testing, and cost about a third as much to run the index. Sonnet 5 wins on hallucination rate (39% vs 88% on AA-Omniscience) and on computer use tooling, and LMArena text is a statistical tie (1461 vs 1466).
The value threat at an introductory $0.75/$3.75 through year-end. It scores higher on the Artificial Analysis Index (41), GPQA (95.3%), MMMU-Pro (85.6%) and LMArena (1493, preliminary), and generates several times faster. Sonnet 5 counters with 128K output against Gemini's 64K, lower hallucination, and a permanent price instead of an introductory one.
Released September 21, 2026 at $2/$6 and scoring 46 on the Artificial Analysis Index at high effort. Sonnet 5 wins on context (1M vs 500K), batch support (Grok 4.7 has none) and output speed at high effort (about 71 vs 54 tokens/sec in Artificial Analysis testing). Grok 4.7's independent results are only days old.
Primary references used to verify this review.
Does not train on API inputs by default
Last verified 2026-09-21