Anthropic Just Made Long Context a Commodity

Anthropic made 1M token context windows generally available for Claude Opus 4.6 and Sonnet 4.6 — and eliminated the long-context surcharge entirely. Previously, any input over 200K tokens cost 2x. Now it’s flat pricing across the full million-token window.

For comparison: Google Gemini and OpenAI GPT-5.4 still charge premiums above 200K tokens. Claude is currently the only model family where both top-tier models offer 1M context at flat rate.

The practical numbers: a 500K-token Opus request that previously cost $5 in input tokens now costs $2.50. The media limit jumped from 100 to 600 images or PDF pages per request. For long coding sessions, fewer forced context compactions — meaning the model retains more of the conversation before needing to summarize and compress.

These are meaningful quality-of-life improvements. But the real story is what flat pricing does to enterprise adoption.

Long context with variable pricing creates a budgeting problem. When costs spike unpredictably based on input length, finance teams impose conservative limits. Developers get told to keep prompts short. Architects design around the cost cliff at 200K tokens instead of using the context window the model actually supports. The surcharge becomes a soft capability ceiling.

Remove the surcharge and the calculus changes:

• Retrieval-augmented generation gets simpler. Instead of complex chunking strategies to stay under 200K, you can stuff more raw context into the prompt and let the model sort it out. Not always ideal, but removes an engineering constraint.
• Document analysis scales. 600 pages per request opens up contract review, compliance checking, and financial analysis workflows that previously required multi-pass architectures.
• Cost forecasting becomes predictable. Per-token pricing without multipliers means usage-based budgets actually work.

The strategic signal here matters. Anthropic is treating long context as table stakes — a baseline expectation, not a premium feature. This is the same trajectory we saw with function calling, vision, and structured output. Features start as differentiators, become standard, then become invisible infrastructure.

If your team has been designing around the 200K context ceiling for cost reasons, it’s time to revisit those architectural decisions. The constraint you were engineering around just disappeared.