Anthropic Drops the Long-Context Surcharge: Claude Opus 4.6 Gets 1M Tokens at Standard Pricing
Anthropic has quietly removed one of the more awkward pricing quirks in its API: the long-context surcharge on Claude Opus 4.6 and Sonnet 4.6. Previously, once a prompt crossed 200,000 tokens, per-token pricing jumped from the standard $5/$25 per million input/output tokens to $10/$37.50. As of Opus 4.6's February 5, 2026 release, that jump is gone. A 1-million-token prompt now costs the same per token as a 10,000-token one.
The 1 million token context window itself launched in beta alongside Opus 4.6, available on the Claude Platform, claude.ai, the Anthropic API, and through AWS, Google Cloud, and Microsoft Azure. It's not available everywhere Claude runs yet, and Anthropic is upfront that it's still a beta feature.
What makes the number meaningful rather than just a marketing figure is how the model performs once that window actually fills up. Anthropic's own MRCR v2 benchmark, which tests whether a model can find and correctly use eight separate pieces of information scattered across a full 1-million-token prompt, put Opus 4.6 at 76%. Sonnet 4.5, the previous generation, scored 18.5% on the same test, evidence that this isn't just a bigger window on the same underlying retrieval quality.
Why It Matters
Long-context pricing has been one of the quiet taxes on building serious AI applications. A surcharge that kicks in once a prompt gets long enough punishes exactly the use cases, large codebases, long documents, extended agent sessions, that benefit most from a bigger context window in the first place. Removing that surcharge doesn't change what the model can do, but it changes who can afford to do it at scale.
For teams already running Opus on large repositories or long documents, this is a direct cost reduction with no code changes required. For teams that were avoiding long-context calls specifically because of the pricing tier, it removes the reason to route around the feature.
Key Details
- Company: Anthropic
- Product: Claude Opus 4.6 and Claude Sonnet 4.6
- Release date: February 5, 2026
- Context window: 1,000,000 tokens (beta, Claude Platform only)
- Maximum output: 128,000 tokens
- Standard pricing: $5 / $25 per million input/output tokens
- Previous long-context surcharge: $10 / $37.50 per million tokens above 200k tokens (now removed)
- Availability: claude.ai, Anthropic API, AWS, Google Cloud, Microsoft Azure
- US-only inference option: 1.1x token pricing multiplier
Technical Analysis
The headline number is the context window: 1 million tokens, roughly the length of a long novel, or a mid-sized codebase. But the number that actually matters more is how the model performs once you fill that window. Anthropic's own benchmark, MRCR v2 with 8 needles at 1 million tokens, tests whether a model can still find and correctly use specific pieces of information buried in a huge prompt, rather than losing track of them, a failure mode researchers call "context rot."
Opus 4.6 scores 76% on that test at the full 1 million tokens. For comparison, Claude Sonnet 4.5 scored 18.5% on the same benchmark, a large gap that suggests the improvement isn't just a bigger window bolted onto the same underlying retrieval quality, but a genuine architecture-level improvement in how the model handles long-range attention.
The tradeoff Anthropic is upfront about: deeper thinking modes (the model has four effort levels, low, medium, high, and max) mean higher latency and cost on simple tasks that don't need it. The 1M window is also still in beta and limited to the Claude Platform rather than being available everywhere Claude runs.
Competitive Comparison
| Product | Better At | Weakness |
|---|---|---|
| Claude Opus 4.6 | Long-context retrieval accuracy (76% on MRCR v2 at 1M tokens), agentic coding (Terminal-Bench 2.0) | Output throughput is slower than Gemini; 1M context still beta, platform-limited |
| Gemini 3.1 Pro | Largest context window (2.5M tokens, verified), fastest throughput (~120 tokens/sec), multimodal benchmarks | Long-context retrieval accuracy trails Claude on MRCR-style needle tests |
| GPT-5.4 / GPT-5.6 | Terminal-Bench 2.0 (75.1%), broad ecosystem and tooling | Smaller standard context window than Gemini; pricing tiers still split by model size |
Industry Impact
Developers building tools that work over large codebases or long documents get a direct cost reduction with zero migration work.
Enterprises running document-heavy workflows, legal review, contract analysis, large-scale codebase audits, can now budget long-context usage at standard rates instead of planning around a premium tier.
Researchers and students working with long papers, datasets, or multi-document analysis benefit from being able to load more material into a single session without hitting a cost wall partway through.
Consumers using claude.ai directly are less affected day to day, since most chat sessions don't approach the 200k-token threshold where the surcharge used to apply.
Future Outlook
The following is analysis and prediction, not confirmed fact.
Pricing moves like this tend to pressure competitors into matching them within a quarter or two, the same pattern played out with token pricing cuts earlier in 2026. If Anthropic's long-context accuracy advantage holds up under independent testing, expect Google and OpenAI to either cut their own long-context pricing or invest visibly in closing the MRCR-style retrieval gap, since "big window, bad retrieval" is now a publicly benchmarked weakness.
The more interesting long-term question is whether context window size keeps mattering as a headline metric at all. Gemini already leads on raw window size at 2.5 million tokens; if the industry converges on "large enough" context windows across the board, the competitive fight likely shifts fully toward retrieval accuracy and cost per useful token, not the size of the number in the marketing copy.
Key Takeaways
- Anthropic removed the pricing surcharge for prompts over 200k tokens on Claude Opus 4.6 and Sonnet 4.6, both now billed at the standard $5/$25 per million token rate.
- The 1-million-token context window is in beta and limited to the Claude Platform.
- Opus 4.6 scores 76% on Anthropic's long-context retrieval benchmark (MRCR v2) at the full 1M tokens, well ahead of Sonnet 4.5's 18.5% on the same test.
- Gemini 3.1 Pro still has the largest raw context window (2.5M tokens) and faster throughput; Claude's edge is retrieval accuracy at long context, not window size.
- This mainly benefits teams already using or considering long-context workloads: large codebases, long documents, and extended agent sessions.
Source: Anthropic
