Claude 4.6 Models Now Feature 1M Context Window at Standard Pricing

Anthropic has launched the general availability of a 1-million-token context window for Claude Opus 4.6 and Sonnet 4.6 at standard pricing. This update increases media limits to 600 items and ensures high recall accuracy across the full window without requiring special beta headers. By eliminating the need for context compaction, the models can now handle massive datasets and long-running agentic tasks more efficiently.

Key Points

Claude Opus 4.6 and Sonnet 4.6 now offer a 1M token context window with no price premium or multiplier.
Media limits have expanded sixfold, supporting up to 600 images or PDF pages per request.
The models maintain industry-leading recall and reasoning performance across the full 1M token window.
General availability includes full rate limits and seamless integration across Claude Platform, AWS, Google Cloud, and Microsoft Azure.
The update eliminates the need for context compaction, allowing AI agents to maintain task adherence and detail over long sessions.

Sentiment

The community is cautiously positive about the 1M context announcement but treats it as an incremental improvement rather than a breakthrough. Most experienced developers emphasize that better context management strategies outweigh raw context size, and the conversation frequently pivots to comparing Claude unfavorably with OpenAI's Codex for handling long sessions. The removal of pricing premiums is appreciated as a competitive response.

In Agreement

The removal of the long-context pricing premium makes the 1M window a genuine competitive advantage, especially for Claude Code users running long autonomous sessions
Larger context windows reduce the frequency of dreaded compaction events, which many developers report as session-destroying — important debugging state gets lost and models loop on already-tried solutions
Enterprise users report the 1M window holding up reasonably well for pair-programming until about 700k tokens, making it useful for sustained development sessions
The expanded media limits (600 images/PDF pages) and no beta header requirement are practical improvements for production API usage

Opposed

Many developers argue that context quality matters far more than context size — focused context under 100k tokens using code maps and auto-context strategies produces better results than dumping everything into a 1M window
OpenAI's Codex reportedly handles compaction far better through encrypted embedding preservation, with some users maintaining coherent sessions spanning billions of tokens — Claude's compaction remains the real bottleneck, not context window size
Multiple users report that the 'dumb zone' of degraded model quality still exists at high context usage, and a larger window just gives more room for the spiral rather than fixing the underlying problem
At Opus 4.6 rates with 700k context, each tool call costs roughly $1 in cache reads alone — 'standard pricing' across 1M tokens is still very expensive for long autonomous sessions
Some developers question whether the benchmark graphs (starting at 256k tokens) conveniently avoid showing the performance degradation curve that users experience below that threshold