Claude Sonnet 5 review: what changed and should you upgrade

Anthropic shipped Claude Sonnet 5 on June 30, 2026, and it’s already the default model behind Claude Free, Pro, Claude Code, and the API. If you’re running any kind of Claude-based coding workflow, the model under the hood changed five days ago whether you noticed or not. Here’s what actually changed, what it costs, and when it’s worth switching versus waiting.

What launched, and when

Claude Sonnet 5 (model ID claude-sonnet-5) replaced Sonnet 4.6 as Anthropic’s default mid-tier model on June 30, 2026. It’s available across Free, Pro, Max, Team, and Enterprise plans, inside Claude Code, through the Claude API, and on both AWS Bedrock and Google Vertex AI.

Anthropic’s pitch is direct: Sonnet 5 performs close to Opus 4.8, at Sonnet prices. That’s a real claim worth checking, not marketing filler, so the rest of this covers the actual numbers.

The model also ships with a 1 million token context window and no long-context pricing surcharge, plus support for up to 128K output tokens. If you’ve been chunking large codebases into smaller Claude Code sessions to stay under context limits, that constraint just loosened considerably.

Pricing: the number that matters and the one that’s hidden

Introductory pricing runs through August 31, 2026: $2 per million input tokens and $10 per million output tokens. After that, standard pricing kicks in at $3 per million input tokens and $15 per million output tokens.

That’s the number every launch article leads with. The one most of them bury or skip entirely: Sonnet 5 uses a new tokenizer, and the same input text can map to 1.0 to 1.35x more tokens than it did under Sonnet 4.6. Anthropic priced the introductory tier to be roughly cost-neutral against Sonnet 4.6 to account for this, which is a reasonable move, but it only holds through the introductory window. Once standard pricing takes over September 1, the token inflation and the price increase stack. A workload that cost $X per run under Sonnet 4.6 could cost meaningfully more than a simple per-token price comparison suggests, especially for prompt-heavy, repetitive work.

If you run scheduled Claude Code pipelines, batch documentation jobs, or any automation that fires the same or similar prompt structure repeatedly, that stacking effect is worth modeling now, before the September price change lands, not after your bill shows up.

The Batch API still offers a 50% discount on both input and output tokens for asynchronous workloads, which is the cleanest lever available if cost is a concern and your workflow doesn’t need real-time responses.

What actually got better

Benchmarks, compared to the outgoing Sonnet 4.6 and the flagship Opus 4.8:

• SWE-bench Verified: Sonnet 5 scores 72.7%, up from Sonnet 4.6’s 62.3%, and closing in on Opus 4.8’s 79.4%.
• SWE-bench Pro (the harder, more realistic variant): Sonnet 5 hits 63.2% against Opus 4.8’s 69.2%.
• FrontierCode: more than doubled versus Sonnet 4.6, jumping from 15.1 to 38.8.

That FrontierCode jump is the most telling number here. It’s a benchmark built around multi-step, real-world coding tasks rather than isolated bug fixes, and doubling that score in one release cycle is a genuine capability shift, not incremental tuning. Anthropic is marketing Sonnet 5 specifically as its most agentic Sonnet yet, built for planning, tool use across browser and terminal, and running multi-step tasks with less hand-holding. The benchmarks back that framing up.

For anyone using Claude Code’s scheduled agents and dynamic workflows, this is the model those pipelines now run on by default, and the agentic gains are the part most directly relevant to that kind of work.

Where it’s not a clean win

Coding generation and agentic task completion are clear improvements. Code review is murkier. Several early comparisons found that Sonnet 5 writes cleaner, more specific review comments than its predecessor, but catches fewer actual bugs than models some teams already have running in production review pipelines, and does it at a slightly higher cost per review pass. If your primary use case is automated PR review rather than code generation, don’t assume the upgrade is automatically better. Test it against your current model on a sample of real pull requests before switching your review pipeline over.

How to actually switch

If you’re on Claude.ai Free or Pro, there’s nothing to do. Sonnet 5 is already the model answering your prompts. If you’re calling the API directly, point your requests at the model ID claude-sonnet-5 instead of the previous claude-sonnet-4-6 string, and check any SDK version pins, since older SDK releases may not recognize the new model ID until updated. Claude Code picks up the new default automatically on update, but if you’ve hardcoded a model flag in a config file or a scheduled task definition, that override will keep running the old model until you edit it directly.

Before flipping a production workload over, run the same prompt through both models and diff the token counts Anthropic reports back in the API response, not just the input text length. That’s the only reliable way to see the tokenizer inflation on your actual prompts rather than trusting the general 1.0 to 1.35x range.

Should you upgrade: a decision framework

Not every Claude workload benefits equally, so the right move depends on what you’re actually running.

Coding agents and autonomous multi-step work. Upgrade now. This is exactly what Sonnet 5 was built for, and the SWE-bench Pro and FrontierCode gains translate directly into fewer stalled agent runs and less manual intervention.

Scheduled automation pipelines (Claude Code cron jobs, n8n workflows calling the API, recurring documentation or content jobs). Upgrade now, but model your token cost first. Run one representative job through both models, compare actual token counts (not just word counts), and confirm the introductory pricing math works for your specific prompts before September 1 changes the equation.

Automated code review. Test before switching. Run Sonnet 5 against your existing model on a batch of real, already-reviewed pull requests and compare bug-catch rates directly. Don’t take either vendor’s framing at face value here.

Casual or one-off use through Claude.ai. Just use it. It’s already the default on Free and Pro, there’s no action required, and the tokenizer cost question doesn’t meaningfully affect a human typing prompts one at a time.

What this means if you’re running IBM i modernization prompts

If you’re using Claude for RPG IV or COBOL documentation work, or running structured prompts against legacy IBM i codebases, the tokenizer change is worth a second look specifically because those prompts tend to be long and repetitive by nature: large code blocks, consistent formatting instructions, the same system prompt fired across dozens of programs. That’s exactly the kind of workload where a 1.0 to 1.35x token inflation compounds fastest. If you’re working from a structured prompt set like the one in the AI Prompt Library or the IBM i Field Guide, it’s worth re-running your token counts against Sonnet 5 before you assume the per-prompt cost is unchanged.

Bottom line

Claude Sonnet 5 is a real upgrade, not a rebrand. The FrontierCode and SWE-bench Pro gains are substantial, the price is genuinely competitive with the previous generation during the introductory window, and if you’re running coding agents or automation pipelines there’s no reason to wait. The catch is the tokenizer, not the model. Do the token math on your actual workloads before the pricing changes September 1, and if automated code review is your main use case, verify the bug-catch rate yourself instead of trusting the launch benchmarks. For everything else, upgrade now and move on.

Frequently Asked Questions

When was Claude Sonnet 5 released?

Anthropic launched Claude Sonnet 5 on June 30, 2026. It replaced Sonnet 4.6 as the default model for Claude Free and Pro plans and is also available on Max, Team, and Enterprise plans, in Claude Code, and through the API.

How much does Claude Sonnet 5 cost?

Introductory API pricing is $2 per million input tokens and $10 per million output tokens, running through August 31, 2026. After that it moves to $3 per million input tokens and $15 per million output tokens.

Is Claude Sonnet 5 better than Claude Opus 4.8?

Not across the board, but close. Sonnet 5 scores 72.7% on SWE-bench Verified against Opus 4.8’s 79.4%, and 63.2% on SWE-bench Pro against Opus 4.8’s 69.2%. Anthropic’s own framing is “close to Opus 4.8 performance, at lower prices,” which is accurate.

Should I switch from Sonnet 4.6 to Sonnet 5 right now?

If you run coding agents, automation pipelines, or anything agentic, yes. If your workload is high-volume batch processing where per-token cost matters more than raw capability, test the new tokenizer’s impact on your specific prompts before committing past September 1.