Claude 1M Context: Anthropic Kills the Long-Context Surcharge

Anthropic has removed the surcharge on million-token context windows — the fee that previously doubled input costs for requests beyond 200,000 tokens. Opus 4.6 and Sonnet 4.6 now price uniformly from 9,000 to 900,000 tokens, with no escalation.

Context

Long-context pricing has been a structural constraint on agentic AI architectures since extended context windows became available. The practical issue wasn’t token limits — it was economic predictability. Workflows that needed to process large codebases, extended documents, or multi-session conversation histories faced a cost cliff: below 200,000 tokens at standard rates, above that at up to double. Building systems that reliably stay under that threshold added architectural complexity without adding capability.

Anthropic’s pricing change removes that cliff entirely. The surcharge is gone. Opus 4.6 and Sonnet 4.6 now operate on a single pricing tier across their full context range, making long-context usage a first-class architectural option rather than a cost-managed exception.

The Detail

The new pricing is straightforward:

• Opus 4.6: $5 per million input tokens / $25 per million output tokens — up to 900,000 tokens per request
• Sonnet 4.6: $3 per million input tokens / $15 per million output tokens — up to 900,000 tokens per request

Previously, requests exceeding 200,000 tokens incurred a surcharge of up to 100% on input pricing. A 500,000-token Opus 4.6 request that previously cost roughly $5 at standard rates would have incurred an additional surcharge bringing the effective rate substantially higher. Under the new structure, the same request costs $2.50 — the standard per-million rate applied uniformly.

The media processing expansion compounds this. Previous limits capped requests at 100 images or PDF pages. That ceiling has moved to 600 — a 6x increase that changes what’s feasible in document-heavy pipelines: legal review, research analysis, financial reporting, and any workflow that requires processing large batches of visual or document-format inputs in a single call.

Both models achieve the highest accuracy among comparable models at full context length in Anthropic’s benchmarks, indicating that performance degradation at maximum context utilization is minimal. The changes apply across Claude’s distribution channels: Claude Code (Max, Team, and Enterprise plans), Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Azure.

Content created with AI assistance and reviewed for accuracy.