Claude 3.5 vs Claude 4: what changed

What actually changed between 3.5 and the 4-series

The current Claude line beats 3.5 at the work most API users actually pay for: agentic coding, multi-step reasoning and staying coherent across long contexts. The API did not change — same Messages endpoint, same request and response shape, same headers — so the interesting question is not "should I switch" but "which ID do I switch to". The answer is below, and the edit itself is one line.

Three improvements are concrete enough to plan around. First, agentic coding: tool use, multi-file edits and long autonomous runs fail noticeably less often than on 3.5 Sonnet, which is why the current models are the default in Claude Code and most coding agents. Second, context: the Opus and Sonnet line serves a 1M-token window at standard pricing, where the 3.5 generation topped out at 200K — long-document and large-repo workloads stop needing chunking workarounds. Third, reasoning control: current models support adaptive thinking with a configurable effort range, so you can buy more deliberation only on the requests that need it.

Output style shifts too. The newer models write denser, more direct prose and follow formatting instructions more literally. That is usually an improvement, but prompts tuned against 3.5's habits deserve a re-run — more on that in the re-testing section.

See also: The best Claude model for coding

Model ID mapping: where each 3.5 model lands

Anthropic retires older model IDs over time, and the current catalog — here and upstream — is the new generation. If your config still names a 3.5-era ID, this is the mapping:

Default to the middle column. The right column exists for one situation: you have prompts or evals pinned to a specific generation and want a proven midpoint while you re-calibrate. Opus 4.7 lists at the same per-token price as Opus 4.8. Sonnet 4.6 is $3/$15 while Sonnet 5 is $2/$10, so staying on 4.6 is a stability choice, not a savings one.

What the upgrade does to your token bill

At list price Sonnet 5 is cheaper than 3.5 Sonnet was: $2/$10 per 1M input/output tokens versus the old $3/$15. Anthropic cancelled the 2026-09-01 flip back to $3/$15. The Opus tier dropped hard: 3 Opus listed at $15/$75, so Opus 4.8 at $5/$25 is a third of what that tier used to cost. Haiku 4.5 lists slightly above the old 3.5 Haiku, but it is a far more capable model for the same slot in your architecture.

On apiToken.sale every request is converted to official Anthropic spend first, then the flat 50% B2C discount is subtracted before it touches your prepaid balance. The ranking between tiers stays the same; every row just lands cheaper than the 3.5-era official bill you were paying.

Estimate your exact workload on the cost calculator →

The migration is a one-line diff

Because the wire protocol is identical, migrating is changing the value of one JSON field. Everything else — endpoint, headers, max_tokens, the messages array, the response's content blocks, stop_reason and usage — stays exactly as your code already handles it.

# Before — Claude 3.5
curl https://router.apitoken.sale/v1/messages \
  -H "x-api-key: sk-pool-•••" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{"model":"claude-3-5-sonnet-20241022","max_tokens":1024,"messages":[{"role":"user","content":"Hello"}]}'

# After — only the model field changes
  -d '{"model":"claude-sonnet-5","max_tokens":1024,"messages":[{"role":"user","content":"Hello"}]}'

On apiToken.sale there is no credential work at all: the same sk-pool key and the same base URL already serve every supported Claude model, so the model ID is genuinely the whole change. If you configure models through an environment variable or a settings panel — ANTHROPIC_MODEL, a Cursor model field, a Continue config entry — update it there and redeploy.

Worth re-testing, not just re-pointing

A protocol-compatible swap is not a behavior-identical swap. Budget one eval pass before you promote the new ID to production:

• System prompts tuned on 3.5: the newer models follow instructions more literally, so workaround phrasing you added for 3.5 ("remember to…", repeated constraints) can now over-constrain the output. Run your prompt suite and delete the scaffolding that is no longer needed.
• Output length: current models tend to answer more thoroughly. If you trimmed max_tokens to keep 3.5 terse, check for stop_reason: max_tokens truncations after the switch.
• Thinking is opt-in: adaptive thinking changes both latency and token spend. Leave it off on latency-sensitive paths until you have measured it, and enable it deliberately on hard-reasoning paths.
• Agent loops: tool-use schemas are unchanged, but the new models call tools more eagerly and recover from tool errors differently. Watch one full agent run before trusting your loop guards.

If a prompt or eval suite is genuinely pinned to older behavior, move to the previous-generation ID (claude-sonnet-4-6 or claude-opus-4-7) as a stepping stone instead of jumping two generations at once — smaller behavioral leap, not a cheaper bill.

Cut over gradually — it is one string per request

There is no account-level migration to schedule, so you can move at whatever granularity suits you: per environment, per feature flag, per request. A common pattern is sending a fixed percentage of traffic to the new model ID while the old config stays intact, then comparing outputs and the dashboard's per-request token detail before committing. Because one key and one prepaid balance cover every supported Claude model — plus GPT, Gemini and Kimi on the same account — a gradual rollout costs you no extra credentials, plans or provider accounts.

Current Claude lineup with per-model pricing →

Frequently asked questions

Related guides