Codex raises Sol’s default context to 500K tokens

• Sol now defaults to 500K tokens in Codex and ChatGPT Work, up from 272K, according to the announcement.
• Subscription-funded usage avoids long-context fees, OpenAI confirmed in its billing follow-up.
• Window size and automatic compaction are independently configurable: MParakhin shared a 1M-window, 900K-compaction setup.

Codex stores personal defaults in ~/.codex/config.toml, including its two context controls. Asana’s browser-agent study found that deleting the previous screenshot on every step broke history-cache reuse.

The increase adds 228K tokens of default headroom, about 84%; OpenAI rounds it to roughly 80%. Its stated goals are fewer compactions and better retention of context and nuance across long-running tasks.

Codex compacts history by summarizing relevant information and discarding less relevant details, according to its manual. Repeated compaction lets it continue working beyond what fits in a single window.

OpenAI had previously defended a smaller default on usage-cost grounds. In an earlier reply, thsottiaux described the Codex setting as “300k” and said running at 1M would consume more usage; he emphasized careful tuning in a later reply.

The fee waiver applies when Sol usage is drawn from a subscription. OpenAI’s follow-up does not announce an API price reduction.

For users who already raised their context window, the fee waiver is the sleeper upgrade. That was theo’s situation: he reported substantial limit-cost savings in his reply.

OpenAI separately lists short- and long-context Standard API pricing in its rate card. Short context means at most 272K input tokens; long context means more than 272K.

Prices per million tokens:

• Input: $2 short context; $4 long context.
• Cached input: $0.10 short context; $0.20 long context.
• Cache writes: $2.50 short context; $5 long context.
• Output: $10 short context; $15 long context.

Codex exposes two independent controls in its configuration reference:

• model_context_window: context-window tokens available to the active model.
• model_auto_compact_token_limit: the token threshold that triggers automatic history compaction. Leaving it unset uses the model default.

MParakhin shared this manual configuration in a reply:

The configurable ceiling remains 1M. No new numerical value for model_auto_compact_token_limit was specified in the announcement.

Caching reuses an unchanged rendered prompt prefix, according to OpenAI’s guide. daniel_mac8 described four factors affecting reuse in his thread:

• History: append-only sessions preserve earlier prompt content instead of rewriting it.
• Settings: changes to the model, reasoning effort, or tools can disrupt matching. The thread exempts GPT-6 Astra from the reasoning-effort restriction.
• Compaction: /compact rewrites earlier context, potentially breaking reuse of conversation history.
• Counters: codex exec --json "<task>" exposes cached_input_tokens and input_tokens; their ratio measures the cached-input share.

Existing cache entries can survive request changes. The guide makes subsequent reuse depend on finding an eligible matching prefix, so a changed request does not automatically erase every reusable entry.

Asana’s optimized Sol browser agent read 89% of its input from cache. daniel_mac8 reported a Sol-only cost reduction from $1.97 to $0.47 per run in his thread.

Asana changed three parts of the workflow in its study:

1. History caching: placed a cache marker on the latest tool result, extending caching beyond tools and the system prompt.
2. Screenshot retention: kept up to 20 screenshots, then pruned back to one. That allowed about 19 consecutive calls to reuse cached history between pruning steps.
3. History budget: increased the budget from 120,000 to 480,000 characters, preventing the earlier text trimming that broke reuse.

The main study comprised 144 runs, with three runs per condition; a 12-run follow-up retained every screenshot. No run reached the 480,000-character budget, and compaction was outside the study’s scope.