Claude Sonnet 4.6 vs Claude Sonnet 5: What Migrating Changes

Anthropic's own documentation currently lists Claude Sonnet 4.6 as both fully active and officially legacy, six weeks after Claude Sonnet 5 replaced it as the default model in Claude Code.

That contradiction sits in two documents published by Anthropic itself. One page marks claude-sonnet-4-6 "Active," with no retirement date scheduled before February 17, 2027. A second page, updated the same week the newer model launched on June 30, 2026, files the older one under a collapsed "Legacy models" accordion and tells developers to consider migrating. For a team running production traffic on the older model today, the real question is not whether it still works. It's what switching actually breaks, and whether this month is the month to do it.

The Verdict: Migrate the Request, Not the Workflow

For most teams building with Claude Sonnet 4.6, swapping the model ID to Claude Sonnet 5 is close to free. Anthropic calls it a drop-in replacement, and for a request with default settings, it is.

The catch sits one layer down. The newer model uses a new tokenizer that turns identical input text into roughly 30% more tokens, and it rejects three request parameters the older one still accepts. A team with tuned max_tokens limits, a temperature-controlled pipeline, or a Priority Tier contract has real engineering work before flipping the switch. A team calling the API with defaults can move today, and probably should.

What Sonnet 4.6 Costs, Right Now

Sonnet 4.6 bills at $3 per million tokens going in and $15 per million tokens coming out, through the Claude API, unchanged since its February 17, 2026 release. It ships with a 1 million token context window, a 128,000-token maximum output, and adaptive thinking that developers opt into with a thinking parameter.

Manual extended thinking, the older budget_tokens-based mode, still runs on this model, though Anthropic marked it deprecated the same week it shipped. It also keeps Priority Tier access, the guaranteed-throughput option Anthropic sells for latency-sensitive production traffic. Its replacement does not have that option at all.

What Sonnet 5 Costs, and Why an Identical Prompt Can Cost More

The newer model launched June 30, 2026 at the same headline rate, with introductory pricing running through August 31, 2026 before standard rates apply. Same eventual numbers, different math underneath them.

The new tokenizer produces about 30% more tokens for identical text, so a request priced a certain way on the older model can cost noticeably more once ported over, even though the rate per token never moved. Anthropic's own migration notes tell developers to recount every prompt with the token-counting API rather than reuse figures measured against the previous generation. Skipping that step is the single most common way a team gets surprised by an invoice.

Neither model changed how billing itself works. What changed is how much of a bill a given sentence generates, and that is a harder thing to notice than a price hike, because nothing on the pricing page moved.

Two Things the Newer Model Took Away

Migrating drops two capabilities outright. Setting an exact budget_tokens value for manual extended thinking returns a 400 error on the newer model. Anthropic wants adaptive thinking and the effort parameter used instead, the same restriction it already applied to Opus 4.7 and Opus 4.8 before this release.

Setting temperature, top_p, or top_k to anything but the default also now fails outright. Neither restriction is new industry behavior; several vendors have moved this direction. It is new for a Sonnet-tier Claude model specifically, and any pipeline built around sampling knobs needs a rewrite, not a config change, before it runs on the newer release.

Where Claude Code Users Already Felt This

The switch is not hypothetical for Claude Code users. Sonnet 5 became the default model for Pro, Team Standard, and Enterprise seats during the week of June 29, 2026, requiring client version 2.1.197 or later.

Anyone who opened the CLI after that week has already been running the newer model by default, changelog read or not. Opus 5 remains the default for Max seats, and Claude Fable 5 sits above both without being anyone's default: Anthropic's own description calls it the most capable widely released model in the lineup, generally available since June 9, 2026.

How the Two Compare on the Numbers That Exist

Anthropic published two hard figures for the older model in the materials accompanying the newer one's launch: 78.5% on OSWorld-Verified, the computer-use benchmark, and 34.6% without tools plus 46.8% with tools on Humanity's Last Exam, a reasoning test.

The comparable scores for the newer model sit inside a chart image in that same announcement rather than in selectable text, so this guide will not repeat numbers pulled from a screenshot as if they were confirmed. Anthropic's own prose describes broad gains in coding, tool use, and agentic tasks, and states the newer model closes distance toward Opus 4.8 on several evaluations. A reader who needs the exact figure for a specific benchmark should open the linked system card directly rather than trust a paraphrase of an image.

For a team still running Claude Sonnet 4.6 in production, that gap is more useful than it sounds. It means nobody, including this guide, can promise an exact percentage-point gain from switching today. What is confirmed is qualitative and directional, not a specific benchmark delta a team can budget a migration around, and pretending otherwise would be inventing precision that does not exist yet.

Where Sonnet 4.6 Still Makes Sense

Three cases argue for staying on the older model past this month. A workload under a Priority Tier contract has no equivalent on the newer release yet. A cost model built tightly around the old token counts needs rebuilding before the same budget holds on the new one, and rebuilding under deadline pressure produces worse decisions than rebuilding on a schedule.

A team that depends on manual extended thinking, setting an exact reasoning budget rather than trusting adaptive judgment, has no migration path until it changes that logic. None of those are reasons to avoid the newer model forever. They are reasons to test before a production cutover, which is not the same thing as waiting indefinitely.

The Turn: Nothing Is Forcing This Migration

The strongest counterargument to all of this is that nothing requires anyone to move. Sonnet 4.6 sits at "Active" status in the official tracker, the same protection Claude Opus 5 and every other current model gets, with a tentative retirement date of not sooner than February 17, 2027.

Compare that to Claude Opus 4.1, which Anthropic deprecated on June 5, 2026 for retirement on August 5, 2026: a hard cutoff, with a named replacement in Claude Opus 4.8. Sonnet 4.6 has received no such notice, on any page, from anyone at the company.

A team with a working, tested pipeline on the older model and no urgent cost or capability problem can reasonably wait. The word "legacy" on the models-overview page describes a product strategy. It is not an operational deadline, and treating it like one is how teams end up rewriting a stable integration for no measurable gain.

What Switching Actually Takes

Anthropic's own migration checklist has three items. Swap the model string from claude-sonnet-4-6 to claude-sonnet-5. Recount token budgets with the token-counting API, because of the new tokenizer. Remove any hardcoded temperature, top_p, top_k, or budget_tokens value before the first production request goes through.

None of the three requires new infrastructure. All three require someone to actually run the numbers instead of assuming the old ones still hold, which is the part most migration guides skip in favor of a one-line "just change the model ID." That line is true and also incomplete.

The safest test is not theoretical. Run a week of real production prompts through both models, compare the resulting token counts and dollar costs directly, and only then decide whether the gains justify the tokenizer's tax on that specific workload. A generic industry benchmark score never tells anyone what their own prompts will cost. Only their own numbers do. That test costs an afternoon of engineering time. Skipping it costs a surprise on next month's invoice instead, and the second one is always more expensive than the first.

The real deadline here is not a retirement notice. It's September 1, 2026, four weeks out from publication, when the newer model's introductory pricing reverts to standard and the token-count increase and the price increase land on the same invoice at once. Nothing about that date is negotiable, and nothing about it requires waiting for Anthropic to send an email first.

Frequently asked questions

Is Claude Sonnet 4.6 deprecated?

No. Anthropic's model-deprecations page lists claude-sonnet-4-6 as Active, with a tentative retirement date of not sooner than February 17, 2027. It has received no deprecation notice, unlike Claude Opus 4.1, which Anthropic deprecated in June 2026 ahead of an August 5, 2026 retirement.

What is the difference between Claude Sonnet 4.6 and Claude Sonnet 5?

Both cost $3 per million input tokens and $15 per million output tokens at standard pricing, with a 1 million token context window. Sonnet 5 uses a new tokenizer that produces about 30% more tokens for the same text, turns on adaptive thinking by default, and removes manual extended thinking and Priority Tier, both of which Sonnet 4.6 still supports.

Is Claude Sonnet 5 the default model in Claude Code now?

Yes. Sonnet 5 became the default model for Claude Code Pro, Team Standard, and Enterprise seats during the week of June 29, 2026, requiring client version 2.1.197 or later. Claude Opus 5 remains the default for Claude Max seats.

Does migrating to Claude Sonnet 5 cost more?

The per-token price is unchanged, but Sonnet 5's new tokenizer means identical prompts produce roughly 30% more tokens, so the same request can cost more even at the same rate. Anthropic's introductory pricing of $2 per million input tokens and $10 per million output tokens runs only through August 31, 2026.

Should I migrate from Claude Sonnet 4.6 to Claude Sonnet 5 right now?

If your integration uses default settings, migrating is close to a one-line change and Anthropic recommends it. If you have tuned token budgets, a Priority Tier contract, or logic built on manual extended thinking, test on Sonnet 5 before cutting over production traffic, since none of those carry over automatically.

Covered in this guide

• Claude Sonnet 4.6: Claude Sonnet 4.6 by Anthropic (Feb 2026) scores 79.6% on SWE-bench Verified with a 1M-token context window at $3/$15 per 1M tokens.
• Claude Sonnet 5: Claude Sonnet 5, released by Anthropic on June 30, 2026, hits 82.1% SWE-bench Verified with a 1M-token context window at $3/$15 per 1M tokens.
• Claude Opus 5: Anthropic's July 2026 flagship LLM, with a 1M token context window by default and a new xhigh reasoning-effort mode for long agentic runs.
• Anthropic: Anthropic, founded 2021 by 7 ex-OpenAI researchers, builds Claude and was valued near $965B after its May 2026 Series H round.
• Claude Code: Claude Code by Anthropic scores 88.6% on SWE-bench Verified, starts at $20/month with no free tier. Reads 1M-token codebases, edits files, runs tests, and opens PRs across 8 platforms.
• Claude Fable 5: The first generally available Mythos-class Claude model, with 1M context and 95.0% SWE-bench Verified. Released June 2026 by Anthropic.
• Claude Opus 4.8: Claude Opus 4.8 (May 2026): Anthropic's most capable model. 1M context window, 88.6% SWE-bench Verified, $5/$25 per 1M tokens. Excels at long-horizon coding and complex reasoning.

Sources

• Models overview
• What's new in Claude Sonnet 5
• Model deprecations
• Introducing Claude Sonnet 5
• Week 27 - June 29 - July 3, 2026