Anthropic just replaced its go-to mid-tier model, and this one changes the math on what “cheap AI model” even means. Claude Sonnet 5 launched on June 30, 2026, and it now runs the free and Pro tiers of claude.ai by default.
Quick answer: Claude Sonnet 5 is Anthropic’s newest mid-tier AI model. It costs $2 per million input tokens and $10 per million output tokens until August 31, 2026, rising to $3/$15 after that. It comes close to matching Anthropic’s more expensive Opus 4.8 model on coding and agent tasks, and it beats GPT-5.5 on most published benchmarks, while costing roughly 40 to 50% less per token.
We tested how it holds up against the claims, checked the pricing details nobody puts in the headline, and compared it against the models people actually ask about — GPT-5.5, Gemini 3.1 Pro, and Anthropic’s own Opus 4.8.
What Is Claude Sonnet 5?
Claude Sonnet 5 is the fifth-generation Sonnet model from Anthropic, sitting between the free-tier Haiku models and the pricier Opus line. The API model ID is claude-sonnet-5.
It ships with a 1-million-token context window and up to 128,000 tokens of output. That puts it in the same league as Gemini 3.1 Pro on raw context size, though Gemini’s output ceiling is smaller at 65,536 tokens. Sonnet 5 also turns on adaptive “thinking” by default and gives you five effort levels to dial that up or down — low, medium, high, xhigh, and max — the same range Opus 4.8 offers.
You can use it right now inside claude.ai (it’s the default model on Free and Pro plans), in Claude Code with /model sonnet, on the Claude Platform, and through Amazon Bedrock.
How Much Does Claude Sonnet 5 Cost?
Here’s where a lot of people get caught out. Anthropic is running introductory pricing of $2 per million input tokens and $10 per million output tokens through August 31, 2026. After that date, it moves to $3/$15 per million tokens, which is the same base rate as Sonnet 4.6.
That sounds like a straightforward win over GPT-5.5’s $5/$30 pricing. It mostly is, but there’s a catch nobody puts on the box: Sonnet 5 uses an updated tokenizer, and the same block of text now produces noticeably more tokens than it used to. Independent testing puts English text at roughly 1.3 to 1.4 times more tokens, code around 1.27 to 1.28 times, and Simplified Chinese barely changed. Artificial Analysis measured the real-world effect directly and found Sonnet 5 cost about 2x more per completed task on their Intelligence Index than Sonnet 4.6 did, and around 15% more than Opus 4.8, before the launch discount is even applied.
So the headline number is real, but your actual bill depends on what you’re feeding the model. Budget for it, especially if your workload is mostly English prose rather than code.
Is Claude Sonnet 5 Actually Better Than Sonnet 4.6?
Yes, and the gap isn’t small. On agentic coding, Sonnet 5 scores 63.2%, up from Sonnet 4.6’s 58.1% — closing in on Opus 4.8’s 69.2%. On the OSWorld benchmark, which tests whether an AI can actually operate a computer (clicking through apps like Google Drive and Excel), Sonnet 5 hit 81.2%, close to human-level performance. For comparison, Claude scored just 28% on the same test back in February 2025, which tells you how fast this specific skill has moved.
On the GDPval-AA v2 knowledge-work benchmark, Sonnet 5 actually edges past the bigger, pricier Opus 4.8, scoring 1,618 versus Opus 4.8’s slightly lower mark. That’s unusual — a mid-tier model outscoring its own flagship sibling on a real task-quality test — and it’s the strongest argument for using Sonnet 5 as your daily driver instead of automatically reaching for Opus.
Where it still trails: pure math reasoning. On the USAMO olympiad-math benchmark, Sonnet 5 scores 79.5% against Opus 4.8’s 96.7% — a 17-point gap that hasn’t closed. If your work leans heavily into formal mathematical proof or olympiad-style logic, Opus 4.8 is still the safer pick.
Claude Sonnet 5 vs GPT-5.5: Who Actually Wins?
On the benchmarks both companies publish, Sonnet 5 comes out ahead more often than not:
| Benchmark | Claude Sonnet 5 | GPT-5.5 |
|---|---|---|
| SWE-bench Pro (coding) | 63.2% | 58.6% |
| Terminal-Bench 2.1 | 80.4% | 78.2% |
| HLE with tools | 57.4% | 52.2% |
| Input price (per 1M tokens) | $3.00 ($2 intro) | $5.00 |
| Output price (per 1M tokens) | $15.00 ($10 intro) | $30.00 |
| ARC-AGI-2 | ~84.7% | 85.0% |
GPT-5.5 keeps a small lead on ARC-AGI-2 abstract reasoning and on long-context retrieval benchmarks like MRCR, where Anthropic hasn’t published a directly comparable number yet. Outside of that, Sonnet 5 wins on coding and tool use while costing meaningfully less per token, even after you account for the tokenizer change. GPT-5.5 still has the bigger production track record, though — it’s been powering OpenAI’s Codex CLI for over two months, while Sonnet 5 launched days ago.
What About Real Work, Not Just Benchmarks?
Benchmarks are one thing. Handing a model your actual messy files is another. Independent testers who ran Sonnet 5 against ordinary office tasks — summarizing email threads, catching planted errors in spreadsheets, briefing long documents — found a genuinely useful pattern: it caught a hidden arithmetic error in a sales table that Sonnet 4.6 missed across four separate test runs. That’s a real jump in how carefully it checks its own work.
It’s not flawless. The same testing found Sonnet 5 tends to write longer than asked, even when told to be brief — about a third more output than GPT-5.5 and roughly double Gemini 3.1 Pro’s length on identical prompts. It also occasionally invented a small unverified detail, like a currency figure that wasn’t in the source material. So it’s more careful with numbers than its predecessor, but it’s still worth double-checking anything it introduces on its own.
On the agent side, early testers at companies like Zapier reported Sonnet 5 finishing multi-step jobs — updating records in one system, then sending a follow-up announcement — that previously stalled halfway through with older models.
Is Claude Sonnet 5 Safe to Use for Real Business Work?
Anthropic’s own safety testing shows Sonnet 5 sitting below Sonnet 4.6 on every measured category of “undesirable behavior” — lower hallucination rate, lower sycophancy, better refusal of malicious requests. In browser-agent scenarios, the success rate of prompt-injection attacks against Sonnet 5 dropped from 31.5% down to 0.93% compared with an unprotected Opus 4.8 setup.
It’s also deliberately weaker on cybersecurity-specific tasks than Opus 4.8 — Anthropic didn’t give it dedicated cyber training, and it comes with default safeguards through what Anthropic calls the Cyber Verification Program. If your work touches security research, expect the occasional extra refusal.
Should You Switch to Claude Sonnet 5?
For most day-to-day coding, content, and agent automation work, yes. It’s a clear step up from Sonnet 4.6, it beats GPT-5.5 on the majority of published benchmarks, and the introductory pricing makes it the cheapest strong option on the market right now through the end of August.
Keep three things in mind before you migrate a production workflow:
- Recount your token usage. The new tokenizer means your real cost per task may run higher than the sticker price suggests, especially for English-heavy prompts.
- Don’t max out the effort setting by default. Pushing effort to “max” roughly doubles cost without a proportional jump in accuracy for most everyday tasks.
- Keep Opus 4.8 on hand for genuinely hard math or research work. Sonnet 5 has closed most of the gap, but not all of it.
For a single very hard task, Opus 4.8 (or the higher-access Fable 5, when available) still earns its price tag. For the dozens of ordinary tasks most teams run every week, Sonnet 5 is now the sensible default.
FAQ SECTION
Q1: What is Claude Sonnet 5? A1: Claude Sonnet 5 is Anthropic’s newest mid-tier AI model, launched June 30, 2026. It’s built for coding, agent automation, and everyday knowledge work, and it’s now the default model on the free and Pro tiers of claude.ai.
Q2: How much does Claude Sonnet 5 cost? A2: Introductory pricing is $2 per million input tokens and $10 per million output tokens through August 31, 2026. After that, standard pricing is $3 per million input tokens and $15 per million output tokens.
Q3: Is Claude Sonnet 5 free? A3: Yes, it’s available for free on claude.ai as the default model on the Free plan, with usage limits. Paid API access uses the token pricing above.
Q4: Is Claude Sonnet 5 better than GPT-5.5? A4: On most published benchmarks — including coding (SWE-bench Pro), terminal use, and tool-assisted reasoning — Sonnet 5 scores higher than GPT-5.5, while costing 40 to 50% less per token. GPT-5.5 keeps a small edge on abstract reasoning (ARC-AGI-2) and long-context retrieval.
Q5: Does Claude Sonnet 5 replace Opus 4.8? A5: For most everyday tasks, yes — it comes close to Opus 4.8’s performance at a much lower price and even beats it on one knowledge-work benchmark. For pure math reasoning and the hardest research tasks, Opus 4.8 still leads by a real margin.
Q6: What is the context window of Claude Sonnet 5? A6: 1 million input tokens, with up to 128,000 tokens of output.
Q7: Why did my API costs go up after switching to Claude Sonnet 5? A7: Sonnet 5 uses an updated tokenizer that turns the same text into more tokens than before — roughly 1.3 to 1.4 times more for English. Recheck your token counts before assuming the lower headline price means a lower bill.
Q8: Where can I use Claude Sonnet 5? A8: On claude.ai (Free, Pro, Max, Team, Enterprise), in Claude Code, through the Claude Platform API, and on Amazon Bedrock.
