The 10 Million Token Window

At Google I/O today, Sundar Pichai announced Gemini 4 with a 10 million token context window. That is roughly 15,000 pages of text, or an entire mid-to-large codebase fed to the model in a single API call. The previous frontier was 2 million tokens (Grok 4.3). Gemini 4 quintupled it in one announcement.

The number sounds transformational. Load your entire codebase. Feed years of conversation history. Give the model everything and let it reason about all of it at once. No chunking, no retrieval augmentation, no information loss from context windowing. Just... everything.

The research says otherwise.

The Breakpoint

Models typically become unreliable 30-40% before their claimed context limit.¹ A 200K model degrades around 130K tokens. A 1M model loses coherence well before the full window. The degradation is not gradual. Researchers describe it as catastrophic, not linear.²

If the 30-40% rule holds for Gemini 4, the effective context window is not 10 million tokens. It is roughly 6 to 7 million. Still enormous. But not 10 million.

The "lost in the middle" problem, documented by Liu et al. in 2023 and confirmed repeatedly since, describes a specific failure mode: models attend to information at the beginning and end of context but lose information placed at 30-70% positional depth, with a 5-15 point retrieval drop in that zone.³ Scaling the window does not fix this. It makes the middle bigger.

The problem with a 10 million token window is not that you can't fit everything in.
The problem is that the model forgets what's in the middle.

The Needle Test

The standard benchmark for context quality is "needle in a haystack": bury a specific fact somewhere in a long context and check whether the model can retrieve it. Simple single-needle tests show high accuracy across frontier models.⁴

Production workloads are not single-needle. They are multi-needle: find the three functions that interact with this database table, identify every place this config value is referenced, trace the call chain from this endpoint to the logging layer. For multi-needle retrieval, effective context sits in the 200K-400K band for most frontier models.⁴ Above that, accuracy degrades meaningfully.

Only Gemini 3 Deep Think held retrieval and reasoning quality through its full window in independent testing.⁴ Whether Gemini 4 inherits that capability is the question Google did not answer today.

The Opus Precedent

We have a recent case study in context window inflation. Claude Opus 4.7 launched in April with a 1 million token context window. Independent testing on the MRCR v2 benchmark showed long-context retrieval dropped from 91.9% to 59.2% compared to Opus 4.6.⁵ A 32.7-point regression confirmed in Anthropic's own system card.

The model got a bigger window. Its ability to use that window got worse. The spec improved. The product regressed.

Google may have solved this. Gemini 3 showed strong long-context performance. But the announcement did not address retrieval quality at the new window size. It announced the size. Size is the easy number to market. Retrieval quality at depth is the hard number to measure.

The Codebase Argument

The developer pitch is: feed your entire codebase into context. No more RAG. No more retrieval pipelines. No more chunking strategies. One codebase, one prompt.

A mid-sized codebase, 500K-1M lines, tokenizes to roughly 2-4 million tokens depending on language and formatting. At 10 million tokens, you could load several codebases simultaneously. The promise is total code awareness, global refactoring, perfect context for every suggestion.

But the research says the effective multi-needle retrieval band is 200K-400K tokens. A 1M-line codebase in a 10M-token window puts most of the code in the zone where retrieval degrades. The model has access to everything. It does not attend to everything equally. The functions buried at token position 4 million are less likely to be found than the functions at position 100K or position 9.9 million.

This is not a theoretical concern. It is a measured phenomenon, reproduced across multiple model families and architectures.³

The Cost

Context is not free. At Gemini 4's expected pricing, a 10 million token input would cost between $10 and $50 per request depending on tier. A refactoring session that makes 20 calls to the full context runs $200-$1,000. Whether the quality at 10M tokens justifies that cost over a well-implemented RAG pipeline at 200K tokens is not obvious.⁶

The industry consensus emerging in 2026 is that context windows will plateau. The focus is shifting to inference-time scaling, better context management, and hybrid approaches that combine compression, caching, and memory-augmented retrieval.⁷ Bigger windows are a brute-force solution. The smart money is on better retrieval.

10 million tokens is a spec sheet number.
The number that matters is how many of those tokens the model actually uses well.

What Google Did Not Say

Google announced the context window size. Google did not announce:

• MRCR or multi-needle retrieval scores at 5M+ tokens
• Lost-in-the-middle benchmarks at the new window size
• Per-request pricing for full-context calls
• Whether the Deep Think retrieval quality from Gemini 3 carries forward
• Latency for 10M token inference

The announcement was the number. The number is impressive. But the number is not the product. The product is retrieval quality, reasoning coherence, and cost at scale. Those numbers will come from independent benchmarks, not keynote slides.

We will test. We will measure. And the effective window will be smaller than 10 million.

It always is.