Claude Sonnet 5: What it does and what it really costs

Anthropic released Claude Sonnet 5 on 30 June 2026, and the timing buried it. The same day, the US government lifted the export controls on Claude Fable 5, and most of the coverage chased the return of the suspended flagship rather than the new mid-tier model. That was a mistake. For most businesses, Sonnet 5 is the more consequential release of the two, because it is the one you will actually use every day.

The pitch is simple enough. Sonnet 5 gets close to the performance of Claude Opus 4.8, then Anthropic’s workhorse flagship, at less than half the price, and at launch on 30 June 2026 it became the default model for everyone on Claude’s Free and Pro plans. The details are where it gets interesting, including a pricing quirk that most of the launch coverage skipped past. Everything below is checked against Anthropic’s announcements and documentation.

Claude model update (checked 7 October 2026)

Anthropic released Claude Sonnet 5.5 on 28 September 2026 at the same $2 and $10 price. Anthropic says it generates output more than 30 per cent faster than Sonnet 5 and costs up to 30 per cent less per task, and Sonnet 5 stays available until at least 30 June 2027. For new work, test Sonnet 5.5 first; our Sonnet 5.5 post covers the effort setting that decides its bill. Anthropic’s Sonnet page now says anyone can chat with Claude using Sonnet 5.5 on Claude.ai, so references below to Sonnet 5 as the Free and Pro default describe the launch; check the model selector for the model your plan uses today. The launch-day evidence below is kept as it was.

What Is Claude Sonnet 5?

Claude Sonnet 5 is the new generation of Anthropic’s mid-tier Sonnet family, released on 30 June 2026. Anthropic calls it the most agentic Sonnet it has built, a claim that needs translating. Give it a job with several steps and it will plan the work, drive tools such as browsers and terminals, check how it is getting on, and see the thing through. Earlier Sonnets would have stopped halfway and asked what comes next.

That word “agentic” is doing real work here. Earlier Sonnet models were quick and cheap but tended to run out of steam on longer tasks, leaving sustained autonomous work to the bigger Opus models. Sonnet 5 narrows that gap considerably. One of Anthropic’s launch testimonials, from an engineer at Zapier, described handing it a two-part job, updating account tiers in Salesforce and then sending a launch announcement to enterprise contacts, and watching it finish end to end where earlier models stalled midway.

It ships with a 1M-token context window as standard, accepts text, images and files, and is available across every Claude plan, in Claude Code, and through the API under the model name claude-sonnet-5. At launch, it replaced Sonnet 4.6 as the default on Claude’s Free and Pro plans. Since 28 September 2026, Anthropic says anyone can chat with Claude using Sonnet 5.5 on Claude.ai, so check the model selector to see which model you are using.

How Good Is It?

Honest answer: very good for the price, and closer to the flagship than any Sonnet before it.

On the agentic coding benchmark Anthropic published, Sonnet 5 scores 63.2% against Opus 4.8’s 69.2%, with its predecessor Sonnet 4.6 back on 58.1%. On one knowledge-work benchmark it actually edges past Opus 4.8, which is the first time a Sonnet has beaten the flagship at anything that matters. The gains over Sonnet 4.6 run across the board: reasoning, tool use, coding and knowledge work all improved, with the system card reporting the full set.

The gap that remains is at the hard end. Opus 4.8 still wins on the most difficult coding and agentic tasks, and Fable 5 sits above both for the long-horizon work that justifies its premium. But the practical shift is that Sonnet 5 and Opus 4.8 now sit on one continuous cost-performance curve rather than two separate tiers with a gulf between them. Anthropic exposes an effort control so you can dial the model’s thinking up or down per task, trading cost against accuracy, and at its highest settings Sonnet 5 reaches into territory that used to require Opus.

There is a safety footnote worth knowing. Sonnet 5 is the first Sonnet-tier model to ship with real-time cybersecurity safeguards, a smaller version of the classifier approach that accompanied Fable 5, and Anthropic reports lower rates of hallucination and sycophancy than Sonnet 4.6. For everyday marketing and business use, neither will get in your way.

What Does Claude Sonnet 5 Cost?

The sticker prices first. Sonnet 5 costs $2 per million input tokens and $10 per million output on the API. Anthropic launched it at those rates as an introductory offer and planned to raise them to $3 and $15 on 1 September 2026, but on 10 August it made $2 and $10 the standard price and said the rise would not happen. That undercuts Sonnet 4.6, which is still listed at $3 and $15. Set it against Opus 4.8 at $5 and $25, Claude Opus 5.5 at $4 and $20, or Fable 5.1 at $10 and $50, and the value case rather makes itself. On subscription plans, Sonnet 5 is simply included. Prices checked on 24 September 2026 against Anthropic’s pricing page and release notes.

Now the catch, and it is the detail most launch coverage missed. Sonnet 5 uses a new tokeniser, the system that chops text into the units you are billed for, and the same text now produces roughly 30% more tokens than it did on Sonnet 4.6, with the exact increase depending on the content. At launch, Anthropic presented the lower price as a way to keep migration roughly cost-neutral, and it has now kept that price. Per-token price parity, in other words, is not per-request price parity. The model also thinks harder by default, which adds tokens of its own.

None of this makes Sonnet 5 expensive. It makes it slightly more expensive than the headline numbers suggest, and it makes the effort dial the real cost control. If you run API workloads, audit your token budgets before switching anything over.

Sonnet 5 or Opus 4.8: Which Should You Use?

Start with Sonnet 5 and escalate only when it fails you. That is the honest rule of thumb now.

For routine content work, analysis, summarisation, coding assistance and most agentic tasks, Sonnet 5 at medium effort delivers what needed Opus a fortnight ago, at a fraction of the cost. Reserve the Opus tier for the genuinely hard problems: the gnarly multi-step reasoning, the highest-stakes work where a wrong answer is expensive, and anything where you have already watched Sonnet 5 fall short. That tier has moved twice since launch. On 24 July Anthropic released Claude Opus 5 at the same $5 and $25 as Opus 4.8, with one hallucination finding worth reading before you switch anything that publishes facts. On 22 September it released Claude Opus 5.5 at $4 and $20, which Anthropic says performs at the level of Claude Fable 5.1 on most work. For new work that outgrows Sonnet 5, Opus 5.5 is now the Opus model to test first. Claude Sonnet 5.5 followed on 28 September 2026 at the same price. We covered where Opus 4.8 earns its keep for marketing work when it launched, and existing 4.8 workflows are under no pressure to move: Anthropic’s deprecation schedule rules out retiring it before late May 2027.

One wrinkle for the perfectionists: at its very highest effort setting, Sonnet 5 can cost more per task than Opus 4.8 running at a comparable accuracy point, because all that extra thinking bills as output tokens. If a task needs the maximum from Sonnet 5, it is worth pricing the same job on Opus before committing. For everything below that ceiling, Sonnet 5 wins the value argument comfortably. On Artificial Analysis’s July benchmark run, even maximum-effort Sonnet 5 came in cheapest per completed task at $1.53, against $1.80 for Opus 4.8 and $2.03 for the newer Opus 5, so the ceiling case is the exception rather than the rule.

What This Means for Marketing Teams

The delegation story we have been tracking all summer just got a price cut. Multi-step marketing work, the competitor round-up assembled from a dozen tabs, the reporting workflow that touches a CRM and a spreadsheet and an email draft, the content audit that used to be a day of clicking, is exactly the shape of task Sonnet 5 was built for, and it now runs at everyday prices rather than flagship ones.

The quieter implication is about your audience. When it launched, Sonnet 5 became the default model for every free Claude user, which meant the answer engines and assistants your customers use to research purchases got noticeably better at reading, comparing and synthesising sources. Content built for extraction, with clear structure, direct answers and verifiable claims, keeps winning those citations. Content that resists machine reading keeps losing ground. The bar rose again on 30 June, for everyone, without anyone opting in.

What This Means for Small Business Owners

If the Fable 5 saga taught one lesson about not depending on a single frontier model, Sonnet 5 teaches the cheerier counterpart: the capability floor keeps rising underneath you at no extra cost. A business in Manchester on a free Claude plan gained, by default, a model that finishes multi-step jobs which stumped the paid flagship of eighteen months ago.

The preparation is unchanged and we will keep repeating it because it keeps being true. Documented processes, structured website content, clean data and clear briefs are what let an autonomous model do useful work for you. The businesses getting real value from these releases are not the ones chasing each launch; they are the ones whose information was in order before the model arrived.

The Price Rise That Did Not Happen

When Sonnet 5 launched, Anthropic said the $2 and $10 API rates would last until 31 August 2026 and then rise to $3 and $15. On 10 August it reversed that: $2 and $10 became the standard price, and the September rise was cancelled. If you built a budget around $3 and $15, you can bring it down. The tokeniser point above still applies, so measure tokens per job rather than assuming the per-token saving carries straight through to the bill.

Its direct successor is now Claude Sonnet 5.5, at the same $2 and $10. When a job outgrows the Sonnet tier, the next step up is Claude Opus 5.5 at $4 and $20. For the longest and hardest work, see Claude Fable 5.1. If you are comparing across vendors, GPT-6 Sol sits at the same $2 and $10 list price on OpenAI’s API.