Every figure, date and benchmark the finished picture puts on screen, chased to a source. Checked 2026-08-31.
GPT 5.1 Codex Max was announced on 19 November 2025 as the first Codex model trained natively to operate across multiple context windows through compaction, coherently working over millions of tokens in a single task. As a session approaches its context limit the model compacts its history, keeping the parts that matter, and continues in a fresh context window.
OpenAI states that in internal evaluations GPT 5.1 Codex Max worked on a single task for more than twenty four hours, persistently iterating on its implementation, fixing test failures and delivering a successful result.
Rendered on screen in beats 040, 041 and 042 as a two column comparison per benchmark.
| Benchmark | GPT 5.1 Codex | GPT 5.1 Codex Max |
|---|---|---|
| SWE bench Verified (extra high reasoning) | 73.7% | 77.9% |
| SWE Lancer IC SWE | 66.3% | 79.9% |
| Terminal Bench 2.0 | 52.8% | 58.1% |
On SWE bench Verified at medium reasoning effort, GPT 5.1 Codex Max outperforms GPT 5.1 Codex at the same reasoning effort while using 30% fewer thinking tokens. Rendered in beat 097.
GPT 5.6 Sol at max reasoning scores 80 on the Artificial Analysis Coding Agent Index, setting a new state of the art and leading all three of the index’s evaluations. It sits 2.8 points above Claude Fable 5, which places Fable 5 at 77.2. Rendered in beat 047.
Artificial Analysis publishes the GPT 5.5 (xhigh) Coding Agent Index score as 76 following the index revision that replaced SWE Bench Pro with Datacurve’s DeepSWE. The narration’s 76.4 is within rounding of the published figure but the extra decimal could not be re-found at the source, so the on screen figure for GPT 5.5 is rendered as 76.
GPT 5.6 Sol reaches 88.8% on Terminal Bench 2.1 and GPT 5.6 Sol Ultra reaches 91.9%. Rendered in beat 048.
OpenAI describes Ultra as a setting that coordinates four agents in parallel by default for harder work. Artificial Analysis compares Ultra’s default four agent configuration against a one agent baseline on BrowseComp, SEC Bench Pro and Terminal Bench 2.1, and reports that adding parallel agents moves the score against latency frontier up and to the left on all three. Rendered in beat 075.
The narration’s pricing figures are the launch rates and are out of date. They are rendered on screen as spoken, so that the picture and the voice agree; the current rates are recorded here and the discrepancy is listed under caveats.
Current published rates, per million tokens:
| Model | Input | Output |
|---|---|---|
| GPT 5.6 Sol | $4.00 | $20.00 |
| GPT 5.6 Terra | $2.00 | $12.00 |
| GPT 5.6 Luna | $0.20 | $1.20 |
Sol launched at $5.00 and $30.00. On 22 August 2026 the standard input, cached input and cache write rates fell 20% and output fell 33.3%, which OpenAI describes as promotional and says will remain available until at least 21 November 2026. Luna’s rates fell by 80% with effect from 30 July 2026.
Anthropic analysed close to four hundred thousand Claude Code sessions recorded between October 2025 and April 2026. Users make about 70% of planning decisions, which answer what to build, while Claude makes about 80% of execution decisions, which answer how to build it. Rendered in beats 052 and 053.
Once Claude Code on the web has started a task the page can be closed and the work continues; when it finishes it opens a pull request against the connected GitHub repository for review. Rendered in beat 055.
Cursor Cloud Agents launched on 24 February 2026 and run in isolated virtual machines, each with a full Linux development environment, terminal, browser and cloned repositories. An agent can select multiple repositories in one environment, make coordinated changes across them, run tests and open pull requests, and it keeps running with the laptop closed. Rendered in beat 057.
The narration’s “more than forty percent” figure is not supported. The picture renders the published figure instead. That figure is 35% of merged internal pull requests as of April 2026, up from 30% at the February 2026 launch. A separate Cursor figure states that sandboxed cloud agents stop 40% less often than unsandboxed ones, which is a different measurement.
Antigravity runs parallel agents and dynamic subagents with scheduled background tasks. The orchestrator monitors every active subagent and surfaces a failure in the terminal, and any subagent’s progress can be opened from the subagent panel. The CLI and the desktop app share the same agent harness. Rendered in beats 060 and 134.
Codex passed five million weekly active users, more than six times its total at the February desktop app launch, with knowledge workers making up roughly 20% of them. Rendered in beat 066.
95% of OpenAI engineers use Codex, and engineers who adopted it open about 70% more pull requests than their peers. This is a company reported internal figure rather than an independent measurement, which the narration says explicitly. Rendered in beat 067.
Codex runs in a sandbox by default. File writes are limited to the workspace and network access is disabled unless the developer enables it. OpenAI recommends reviewing the agent’s work before deploying it. Rendered in beats 106 and 107.
Figures that were recorded before they could be checked, and how the picture handles them.