Pricing, Limits, Lifecycle and Compliance
Production and enterprise interviews ask these practical details: the cost of tools, tokens for images, rate limit tiers, when a model is retired, and compliance certifications.
27.1 Pricing of tools and features
| Item | Price |
|---|---|
| Web search | $10 per 1,000 searches + tokens for the results (no charge on error) |
| Web fetch | No extra charge, only tokens for the fetched content |
| Code execution | Free with web search/fetch (newer versions); otherwise 1,550 free hours/month/org, then $0.05 per container-hour |
| Managed Agents session | Tokens + $0.08 per session-hour (running time only) |
| Batch API | 50% off input and output |
| Cache write / read | 1.25x (5 min) / 2x (1 hour) write; read ~0.1x (Opus 5.5: $0.20/MTok) |
| inference_geo: "us" | 1.1x multiplier (Claude 4.6+); global default is standard price |
| Long context | 4.6+ models (except Haiku 5.5) get the full 1M context at standard price; Haiku 5.5 has a higher price on 100K+ prompts |
| Computer use toolset | ~4,500 extra input tokens per request (tool definitions) |
| Free credits | New API users get a few free credits to test |
27.2 Tokens for images
An image is processed in 28×28 pixel patches: tokens = ⌈width/28⌉ × ⌈height/28⌉. Large images are downscaled.
| Point | Value |
|---|---|
| High-res tier (Claude 4.7+) | Max long edge 2,576 px, max 4,784 visual tokens |
| Standard tier (older models) | Max 1,568 px, 1,568 tokens |
| 1000×1000 image | ~1,296 tokens (~$5 per 1,000 images on Opus 5.5) |
| Max per request | 600 images (200K context models: 100); claude.ai: 20 per message |
| Max size | 8000×8000 px; 10 MB per image (Bedrock 5 MB); stricter limit with 20+ images (~2000 px) |
| Formats | JPEG, PNG, GIF (first frame only), WebP |
| Limitations | Doesn't identify people, counting is approximate, can't detect AI-generated images, not for CT/MRI diagnosis |
27.3 Rate limits and usage tiers
| Tier | Monthly spend cap | Opus 5.5 RPM / ITPM / OTPM |
|---|---|---|
| Start | $500 | 1,000 / 2M / 400K |
| Build | $1,000 | 5,000 / 5M / 1M |
| Scale | $200,000 | 10,000 / 10M / 2M |
| Custom | No cap (via sales) | Negotiated |
- Token bucket algorithm: capacity refills continuously, with no fixed reset. 60 RPM can also be enforced as ~1 request/second.
- Cache-aware ITPM: on most models cache_read_input_tokens don't count toward ITPM, so caching multiplies effective throughput.
- OTPM does not count max_tokens, only the tokens actually generated.
- Hitting the spend cap = 429 enforced_spend_limit_reached (without retry-after): retrying is useless, raise your tier.
- Response headers: anthropic-ratelimit-requests-remaining, ...-input-tokens-remaining, retry-after.
- You can set your own lower limits on Workspaces; Batch, Files and Managed Agents have separate limits.
27.4 Model lifecycle
| Status | Meaning |
|---|---|
| Active | Fully supported, recommended |
| Legacy | No more updates, may be deprecated in the future |
| Deprecated | Still works but not recommended; replacement + retirement date announced |
| Retired | Requests fail |
| Recent retirements | Retire date | Replacement |
|---|---|---|
| Claude Sonnet 4.5 | 30 Nov 2026 (deprecated) | claude-sonnet-5-5 |
| Claude Opus 4.1 | 5 Aug 2026 | claude-opus-4-8 |
| Claude Opus 4 / Sonnet 4 | 15 Jun 2026 | claude-opus-4-8 / claude-sonnet-5-5 |
| Claude Haiku 3 | 20 Apr 2026 | claude-haiku-4-5-20251001 |
| Claude Sonnet 3.7, Haiku 3.5 | 19 Feb 2026 | claude-sonnet-5-5 / Haiku 4.5 |
- For public models, an email notice at least 60 days before retirement.
- Use Console Usage → Export to see which API key is using which old model.
- Anthropic has committed to preserving the weights of retired models.
- Bedrock / Google-managed Vertex keep their own separate retirement dates.
27.5 API versioning and beta headers
| Concept | Detail |
|---|---|
| anthropic-version header | The API version (like 2023-06-01, newer 2026-10-21). An SDK major version sends one API version |
| 2026-10-21 change | temperature, top_p, top_k removed: sending them gives a 400; control behavior through prompting |
| anthropic-beta header | Opt in to beta features (like compact-2026-01-12); in the SDK betas=[...] + client.beta.* |
| Beta → GA | Once a feature is stable the header is no longer needed (like the Files API, Skills) |
| Migration guides | Breaking changes of each new model (prefill removed, budget_tokens removed...) |
27.6 Compliance and data
| Certification / Program | Detail |
|---|---|
| SOC 2 Type I and Type II | Audit of security controls |
| ISO 27001:2022 | Information Security Management |
| ISO/IEC 42001:2023 | AI Management Systems (AI-specific standard) |
| HIPAA | HIPAA-ready configuration, BAA available (commercial) |
| Trust Portal | trust.anthropic.com: request reports and documents |
- Commercial (API, Team, Enterprise) data is not used for model training by default.
- Zero Data Retention (ZDR) is for eligible customers; some models (Fable) require 30-day retention.
- Data residency: inference_geo, cloud regions (Bedrock/Vertex/Foundry), workspace-level controls.