Codex GPT-5.6 Sol context window reduced to 258K tokens, users report performance degradation

What the GPT-5.6 Sol context window is

A context window is the maximum amount of text a model can consider at once, measured in tokens: the prompt, the conversation so far, any files or documents loaded in, and the response being generated all have to fit inside it. When a session outgrows the window, something has to be dropped or compressed, and the model loses direct access to whatever fell out.

GPT-5.6 Sol is the variant of GPT-5.6 that OpenAI runs in Codex, its agentic coding product. OpenAI advertises a context window of up to 1.05 million tokens for the model. That figure describes the model's ceiling, not what any given product serves: the window a Codex session actually gets has consistently been smaller, and it is the gap between those two numbers that this story is about.

What changed in Codex

In July 2026, users measuring their Codex sessions found the effective context window had dropped from roughly 353,000 tokens to roughly 258,000 — a cut of about 27 percent, made without a headline announcement. The same reports describe severe slowdowns, with sessions that previously ran comfortably now degrading well before the old limit.

For a coding agent, the practical difference is larger than the raw numbers suggest. Codex sessions accumulate context quickly: repository files, diffs, test output, and the agent's own working notes all compete for the same budget. A window cut of nearly 100,000 tokens means long refactors and large-repository work hit truncation sooner, and the agent starts forgetting earlier parts of the task it is still in the middle of.

Why the effective window is smaller than the advertised limit

The 1.05 million token figure and the 258K figure are both real; they just measure different things. The advertised number is the maximum the model architecture supports, typically via the raw API at full price. The effective number is what a product chooses to serve, and products almost always serve less.

There are three standard reasons. Cost: serving long contexts is expensive, and inference cost grows faster than linearly with context length, so capping the window is the most direct lever a provider has on margins. Quality: models degrade on very long contexts, with retrieval accuracy in the middle of the window falling off well before the hard limit, so a smaller window can genuinely produce better answers. And overhead: an agentic product spends a meaningful slice of the window on its own system prompts, tool definitions, and scaffolding before the user's content gets a single token.

None of that is unique to OpenAI or to Codex. What drew attention here is the direction and the silence: an already-reduced effective window was reduced further, discovered by users measuring it rather than announced.

What it means for developers using Codex

Sessions truncate sooner. Work that spans many files, long test logs, or extended back-and-forth will cross the 258K boundary earlier than before, and once compaction kicks in the agent is working from summaries of its earlier context rather than the context itself. Symptoms include the agent re-reading files it already read, forgetting decisions made earlier in the session, and quality dropping late in long tasks.

The reported slowdowns compound this. A slower session encourages leaving one long-running conversation open rather than starting fresh ones, which is exactly the usage pattern the smaller window now punishes.

How to work within the smaller window

Scope sessions to one task. The context budget resets with each new session, so a fresh session per task keeps the whole window available for the work at hand instead of the residue of the last three tasks.

Keep the working set small. Point the agent at the specific files and directories that matter rather than letting it ingest broadly; every file it reads is spent budget. Externalise durable state: decisions, task lists, and specs survive better in a file the agent re-reads on demand than in conversation history that will be compacted away.

Summarise before you continue. When a long session is unavoidable, asking the agent to write its own handover summary and starting a new session from that summary preserves the decisions while releasing the tokens. And if a task genuinely needs more simultaneous context than Codex now serves, the raw API with an explicit long-context configuration remains the escape hatch, at raw-API prices.