TokenMix Research Lab · 2026-04-24

Claude 3.7 Sonnet Pricing 2026: Retired, Upgrade To 4.6
Last Updated: 2026-04-30
Author: TokenMix Research Lab
Data checked: 2026-04-30
Claude 3.7 Sonnet is retired on the Claude API. Do not price new workloads on 3.7. Upgrade to Claude Sonnet 4.6 unless you have a specific 4.5 regression reason.
The old 3.7 pricing question used to matter because Claude 3.7 Sonnet introduced hybrid reasoning at $3/$15 per million tokens. But Anthropic's current model deprecations page lists claude-3-7-sonnet-20250219 as retired, with retirement on February 19, 2026. It also lists claude-sonnet-4-6 as the recommended replacement for retired Sonnet 3.7 and older Sonnet models. So the real 2026 answer is not "is 3.7 still worth it?" It is "how do I migrate with the least quality, cost, and behavior risk?"
Table of Contents
- Quick Verdict
- Confirmed Status
- Historical 3.7 Pricing vs Current Sonnet Pricing
- Upgrade Target: 4.6 vs 4.5 vs Opus
- Migration Steps
- Extended Thinking Migration
- Cost Math
- When 4.5 Still Makes Sense
- What Not To Do
- Final Recommendation
- FAQ
- Related Articles
- Sources
Quick Verdict
Do not start any new work on Claude 3.7 Sonnet. It is retired. Migrate old API calls to Sonnet 4.6, then reserve Opus 4.7 only for tasks where Sonnet fails.
| Situation | Best action | Why |
|---|---|---|
| New app | Use claude-sonnet-4-6 |
Current recommended Sonnet route |
| Old 3.7 API code | Migrate to 4.6 | 3.7 is retired |
| Need exact old behavior for audit | Keep old outputs, not old API dependency | Retired model is not a stable production path |
| Worried 4.6 changes style | A/B against 4.5 and 4.6 | 4.5 is active longer than 3.7, but 4.6 has longer runway |
| Hard reasoning/coding still fails on Sonnet | Route to Opus 4.7 | Higher cost, higher capability |
| Cost-sensitive simple tasks | Route to Haiku 4.5 or cheaper model | Sonnet is not needed for every step |
Confirmed Status
| Model | Current API state | Tentative retirement signal | Recommended reading |
|---|---|---|---|
claude-3-7-sonnet-20250219 |
Retired | Retired February 19, 2026 | Do not use for new or production API calls |
claude-sonnet-4-20250514 |
Deprecated | Retires June 15, 2026 | Migrate before the cutoff |
claude-sonnet-4-5-20250929 |
Active | Not sooner than September 29, 2026 | Use for regression only if needed |
claude-sonnet-4-6 |
Active | Not sooner than February 17, 2027 | Default Sonnet upgrade target |
claude-opus-4-7 |
Active | Not sooner than April 16, 2027 | Premium fallback for hardest tasks |
claude-haiku-4-5-20251001 |
Active | Not sooner than October 15, 2026 | Low-cost route for simple tasks |
This table should replace any old rule of thumb about "18-month support" or "safe through 2027." Use the official deprecation page, not lifecycle guesses.
Historical 3.7 Pricing vs Current Sonnet Pricing
Claude 3.7 Sonnet launched with the same Sonnet price band that later Sonnet releases kept: $3 per million input tokens and $15 per million output tokens. That historical fact is useful for migration math, but not for choosing 3.7 today.
| Model | Price signal | Current status | What it means now |
|---|---|---|---|
| Claude 3.7 Sonnet | Historical $3/$15 per MTok | Retired | Do not budget new traffic here |
| Claude Sonnet 4.5 | $3/$15 per MTok | Active | Shorter runway than 4.6 |
| Claude Sonnet 4.6 | $3/$15 per MTok | Active | Best default upgrade |
| Claude Haiku 4.5 | $1/$5 per MTok | Active | Cheaper for simple work |
| Claude Opus 4.7 | $5/$25 per MTok | Active | Higher-cost premium fallback |
The important point: moving from 3.7 to Sonnet 4.6 does not change the headline Sonnet price band. It changes model behavior, support runway, and available features.
Upgrade Target: 4.6 vs 4.5 vs Opus
The correct replacement is usually Sonnet 4.6. Sonnet 4.5 is a temporary compatibility target. Opus 4.7 is a premium escalation target, not the default replacement for every 3.7 call.
| Target | Use when | Avoid when |
|---|---|---|
| Sonnet 4.6 | New default, coding, agents, analysis, app backends | You need exact 4.5 behavior for a known regression |
| Sonnet 4.5 | Regression shows 4.6 changes output in a harmful way | New projects or long runway needs |
| Opus 4.7 | Hard reasoning, difficult code review, high-value agent tasks | Classification, extraction, routine chat |
| Haiku 4.5 | Fast low-cost routing, pre-processing, simple extraction | Deep reasoning or complex code generation |
| Multi-model gateway | You need fallback, cost routing, or A/B testing | You only need one fixed first-party model |
For broader routing logic, see our Claude Sonnet vs Opus guide and Claude Haiku vs Sonnet guide.
Migration Steps
Do not migrate by changing a model name and hoping. Retired-model migrations deserve a small regression set.
| Step | Action | Why |
|---|---|---|
| 1 | Search code and config for claude-3-7-sonnet-20250219 and aliases |
Find every old route |
| 2 | Capture 20 to 50 real prompts | Generic prompts miss production drift |
| 3 | Run Sonnet 4.6 against the set | Establish replacement behavior |
| 4 | Compare task success, not vibes | The model can sound different while solving better |
| 5 | Check output length and token cost | Same price band can still produce different token usage |
| 6 | Update fallback chains | Avoid falling back to retired or deprecated models |
| 7 | Monitor 429/529 and latency | Production behavior matters |
Minimal code change:
from anthropic import Anthropic
client = Anthropic(api_key="your-anthropic-api-key")
message = client.messages.create(
model="claude-sonnet-4-6",
max_tokens=800,
messages=[{"role": "user", "content": "Run the migrated prompt."}],
)
print(message.content[0].text)
If you use a gateway such as TokenMix.ai, run the same regression set through Sonnet 4.6, Sonnet 4.5, Haiku 4.5, Opus 4.7, and non-Claude fallback models. The right default is the model with the lowest cost per successful task, not necessarily the model with the highest benchmark score.
Extended Thinking Migration
Claude 3.7 Sonnet mattered historically because it introduced hybrid reasoning: normal answers or extended thinking with a controllable thinking budget. Newer Claude models also support thinking modes, but you should still test cost and output behavior.
| 3.7-era behavior | Migration check |
|---|---|
| Thinking tokens billed as output | Confirm current model pricing and output budget |
| Large thinking budgets for math/code | Re-test whether Sonnet 4.6 needs the same budget |
| Long reasoning chains | Watch latency and OTPM pressure |
| Tool-heavy agent prompts | Re-test tool choice and stop conditions |
| Prompt tuned for 3.7's style | Remove unnecessary scaffolding if 4.6 follows better |
Do not blindly carry over huge thinking budgets. If Sonnet 4.6 solves the task with less reasoning budget, keep the savings.
Cost Math
Assume 100 million input tokens and 30 million output tokens per month.
| Route | Input cost | Output cost | Monthly cost | Notes |
|---|---|---|---|---|
| Historical 3.7 Sonnet band | $300 | $450 | $750 | Retired; use only as baseline |
| Sonnet 4.6 | $300 | $450 | $750 | Default replacement |
| Sonnet 4.5 | $300 | $450 | $750 | Active but shorter runway |
| Haiku 4.5 | $100 | $150 | $250 | Good for simple tasks |
| Opus 4.7 | $500 | $750 | $1,250 | Premium escalation |
| 70% Haiku + 25% Sonnet + 5% Opus | Mixed | Mixed | About $425 | Routing can beat one-model defaults |
The migration opportunity is not "save money by keeping 3.7." That path is gone. The opportunity is to route tasks more intelligently after migration.
When 4.5 Still Makes Sense
There are narrow cases where 4.5 is a better intermediate step than 4.6.
| Case | Why 4.5 may help |
|---|---|
| 4.6 changes a strict output format | 4.5 may be closer to old behavior |
| Creative style regression | Some workflows value old tone |
| Compliance validation takes months | 4.5 is active and buys time |
| Historical benchmark continuity | 4.5 is closer to the old generation than 4.6 |
| Gateway currently exposes 4.5 but not 4.6 | Temporary route only |
But for new work, 4.5 should not become the new long-term default. It has a shorter listed support runway than 4.6.
What Not To Do
| Bad idea | Why |
|---|---|
| Keep recommending 3.7 as stable | It is retired on the Claude API |
| Compare 3.7 benchmarks as if it is buyable today | Historical data does not equal current availability |
| Migrate to 4.5 by default | 4.6 is the recommended replacement for retired Sonnet 3.7 |
| Ignore deprecation docs | Blog guesses age badly |
| Route fallback to Sonnet 4 | Sonnet 4 is already deprecated with a June 15, 2026 retirement date |
| Treat Opus as a drop-in replacement for all 3.7 traffic | It is materially more expensive |
Final Recommendation
Claude 3.7 Sonnet pricing is now historical. Replace 3.7 with Sonnet 4.6, keep Sonnet 4.5 only for regression exceptions, and use TokenMix.ai routing to send easy work to cheaper models.
FAQ
Is Claude 3.7 Sonnet deprecated or retired?
It is retired on the Claude API. Anthropic's model deprecation page lists claude-3-7-sonnet-20250219 as retired, with retirement on February 19, 2026.
What should I replace Claude 3.7 Sonnet with?
Use claude-sonnet-4-6 for most workloads. Anthropic lists Sonnet 4.6 as the recommended replacement for retired Sonnet 3.7 and older Sonnet models.
Was Claude 3.7 Sonnet priced at $3/$15 per million tokens?
Yes historically. Anthropic's 3.7 launch said it had the same price as its predecessors: $3 per million input tokens and $15 per million output tokens, including thinking tokens.
Does moving to Sonnet 4.6 increase headline price?
No, not at the starting Sonnet API price band. Sonnet 4.6 is also priced at $3/$15 per million tokens. Output length and prompt behavior can still change effective cost.
Should I migrate to Sonnet 4.5 first?
Usually no. Use Sonnet 4.6 unless your regression tests show a specific 4.6 behavior problem. Sonnet 4.5 has a shorter listed support runway.
Does extended thinking still matter after 3.7?
Yes, but re-test budgets. Newer Claude models may solve some tasks with different thinking settings, and large reasoning budgets can raise output cost and latency.
Can TokenMix.ai help with 3.7 migration?
Yes. Use TokenMix.ai to A/B Sonnet 4.6, Sonnet 4.5, Haiku, Opus, and non-Claude fallback models through one API surface, then route by task type and cost per successful output.
What if my app requires exact 3.7 outputs?
Archive old outputs for audit, but do not rely on retired live API availability. Build a regression suite and choose the closest active replacement.
Related Articles
- Claude Sonnet 4.5 Free Access 2026: API Test vs 4.6 Safely
- Claude Sonnet 4.6 Free Trial 2026: 5 Safe API Test Paths
- Claude API Pricing 2026: Opus, Sonnet, Haiku Costs Compared
- Anthropic API Pricing 2026: Cache, Batch, Data Residency Fees
- Claude Sonnet vs Opus 2026: Pricing, Quality, Routing Guide
- Claude Haiku vs Sonnet 2026: Cost, Quality, Routing Rules
- AI API Gateway 2026: 7 LLM Routing and Fallback Options