Anthropic released Claude Sonnet 5 on July 1, 2026, and it changes the calculus for any business running AI agents. For the first time, you get near-Opus performance at Sonnet pricing — and with introductory rates of $2 per million input tokens through August 31, the timing is deliberate. Enterprises that burned through budgets on agentic workloads in Q2 now have a path back.

What Actually Changed

The biggest shift isn’t a capability number. It’s that Claude Sonnet 5 ships with adaptive thinking on by default. Every request now runs with the model’s reasoning process active, without you needing to configure anything. That’s the same mode that made Opus 4.8 so effective on complex tasks — now it’s the baseline for Sonnet.

That changes the profile of what Sonnet can handle. Agentic workflows that previously required bumping to an Opus-class model — multi-step research, tool use across browsers and terminals, long-horizon planning — are now in Sonnet territory.

Other notable changes in the release:

• 1 million token context window by default, supporting very large documents, full codebases, and long conversation histories
• 128,000 max output tokens, making it suitable for lengthy generation tasks without chunking
• Cybersecurity safeguards enabled by default — the first Sonnet-tier model to carry them. Requests touching prohibited or high-risk cybersecurity topics return a stop_reason: "refusal" in a normal HTTP 200 response rather than an error, so existing API consumers won’t see hard failures
• New tokenizer that generates roughly 30% more tokens for the same input text compared to Sonnet 4.6, which affects budgeting even though per-token pricing is unchanged

The tokenizer change is worth calling out plainly: if you migrate from Sonnet 4.6 and reuse your token budgets or cost estimates, they will be off. The introductory pricing partially offsets this, but you should recount your typical prompts before committing to production rollout.

The Pricing Picture

The introductory rate runs through August 31, 2026:

	Input	Output
Introductory (now–Aug 31)	$2/M tokens	$10/M tokens
Standard (from Sep 1)	$3/M tokens	$15/M tokens

Standard pricing matches what Sonnet 4.6 cost — but with the new tokenizer producing 30% more tokens for the same content, equivalent requests will cost roughly 30% more at standard rates than they did on Sonnet 4.6. Worth factoring into any budget projections beyond August.

That said, the capability-per-dollar improvement is real. Workloads that previously required Opus 4.7 or 4.8 can move to Sonnet 5, and those models cost $15/$75 per million tokens at standard rates. If your use case doesn’t need the absolute top of the range, Sonnet 5 is a meaningful step down in cost.

Where It’s Available

Claude Sonnet 5 is available immediately across:

• Anthropic API — all customers
• Amazon Bedrock — via Claude in Amazon Bedrock and Claude Platform on AWS (not available on the legacy InvokeModel/Converse APIs)
• Google Cloud Vertex AI — available now
• Microsoft Foundry — in preview

On Claude.ai, it becomes the default model for Free and Pro subscribers. Max, Team, and Enterprise plans have access alongside existing models.

Migration: Mostly Straightforward

Anthropic positioned this as a drop-in upgrade. The model ID change is claude-sonnet-4-6 to claude-sonnet-5. Beyond that, three things need review:

1. Remove manual extended thinking. Setting thinking: {type: "enabled", budget_tokens: N} now returns a 400 error. Use adaptive thinking with the effort parameter instead.
2. Remove custom sampling parameters. Setting temperature, top_p, or top_k to non-default values returns a 400 error, matching behavior already in place for Opus-class models.
3. Recheck token budgets. Any max_tokens limit tuned close to expected output length on Sonnet 4.6 may now truncate output on Sonnet 5 due to the new tokenizer.

If your code was already running cleanly on Sonnet 4.6, the migration is a one-line change — unless you relied on the two deprecated parameters.

What This Means for Business

The timing of this release is worth reading in context. Anthropic is heading toward an IPO, the AI agent cost crisis of Q2 2026 left real bruises on enterprise budgets, and the Fable 5 export restrictions created uncertainty for some international operations. Sonnet 5 is the answer to most of those problems at once: it’s capable enough to handle serious agentic work, widely available, and cheaper than the alternatives at comparable performance.

For businesses running AI agents — customer support, internal knowledge retrieval, data analysis workflows, document processing — Sonnet 5 is probably the right default choice now. It covers the middle ground that most enterprise AI workloads actually need, without paying Opus rates.

The bigger picture is that the performance gap between model tiers keeps narrowing. Sonnet 5 performing close to Opus 4.8 means the “upgrade to Opus when it gets hard” heuristic needs revisiting. For most structured business tasks, you may not need to go there at all.

What to Watch

The tokenizer change and the August 31 introductory pricing cutoff are the two items to track. If you’re building agent pipelines now, size them against September pricing, not July pricing — that’s the sustainable number. And benchmark your actual prompts against the new tokenizer before making cost commitments to stakeholders.

---

If you’re building with Claude or Codex right now, grab the free Working With Claude field guide. Thirty-two pages on the full ecosystem, Claude Code in depth, and how to roll agents out properly. Get the free guide.