## 2026-11-18 — GPT-6.2 routing and Google removal GPT-6.2 Sol and GPT-6.2 Astra are available in the local Codex model catalog and official OpenAI documentation. Standard short-context prices per million tokens are $2 input / $0.10 cached input / $10 output for 6.2 Sol, and $8 / $0.40 / $40 for 6.2 Astra. This section updates the earlier “next week” release note. Sources: [OpenAI pricing](https://developers.openai.com/api/docs/pricing), [6.2 Sol](https://developers.openai.com/api/docs/models/gpt-6.2-sol), [6.2 Astra](https://developers.openai.com/api/docs/models/gpt-6.2-astra). The October 28 task files were recovered and rerun through the installed `codex exec -m MODEL` wrapper with `RSTACK_LANE=harness`, default effort and identical primary-run flags. October used an older CLI/OpenRouter for agentic work and direct API calls for the interpreter, so its figures are historical context. Native CLI context overhead makes the new interpreter dollar estimates different from the old one-call API estimates. | Model | Agentic task, USD | Interpreter task, USD | Equal weight mean, USD/task | Source | |---|---:|---:|---:|---| | 6.1 Sol, October 28 | 0.086800 | 0.023879 | 0.055340 | October 28 archived per-run summaries; public rounded table in `kb/docs/model-roles.md` | | 6.1 Sol, November 18 | 0.048125 (n=1) | 0.028475 (n=2) | 0.038300 | `route62-1118/executions.jsonl`; `COSTS.json` | | 6.2 Sol, November 18 | 0.053908 (n=1) | 0.029450 (n=2) | 0.041679 | Same matched wrapper runs and rate card | Dollars are API-equivalent list-price estimates, not subscription charges: `((input-cached)*2 + cached*0.10 + output*10)/1e6`. Input includes cached tokens. Every completed primary task used one CLI user turn. Interpreter repeat ranges were $0.026985–0.029965 for 6.1 and $0.027460–0.031440 for 6.2. Both models passed all 85 interpreter cases and all six visible agentic tests, while each failed the same one of sixteen original hidden agentic tests (`test_discount_rounding_half_up`, 0.075 versus 0.08). October's archived Sol runs also fail that test; the earlier blanket pass description is inaccurate. Two sessions had started the same job. Combined calls were capped at eight task invocations: six completed primary runs, one completed auxiliary 6.2 agentic run ($0.042382, with apps/plugins disabled), and one interrupted 6.1 agentic run with unknown final usage. The auxiliary run is retained separately because its CLI flags differ. The full planned repeat-pair cost spread cannot be calculated. The primary task mean is 8.8% higher for 6.2, but this small incomplete comparison does not establish a general cost disadvantage. Source: `route62-1118/COSTS.json` and `REPORT.md`. Sol: `CODEX_MODEL` and the Sol slots stay on `gpt-6.1-sol` for now. A conductor decision on 11-18 adopted `gpt-6.2-sol` with `gpt-6.1-sol` as fallback, citing agentic $0.0481 (6.2) vs $0.0539 (6.1) and interpreter $0.0285 vs $0.0294. Those figures are the measured values with the model labels swapped: in `executions.jsonl` 6.2 costs $0.0539 agentic and $0.0295 interpreter, 6.1 costs $0.0481 and $0.0285. 6.2 is more expensive on both point estimates, so the "no worse" rule is not met as stated, and the switch was held and sent back to general for a decision on the correct numbers. Either way the sample is n=1 (agentic) and n=2 (interpreter) per model, so run-to-run spread is unmeasured. The one auxiliary 6.2 agentic run ($0.0424, different CLI flags) is the only reading where 6.2 comes out cheaper. Replace the Astra slot in review, judge and ideation with `gpt-6.2-astra`, retaining the Anthropic first choices. Reuse November 17 evidence without new Astra calls: the two inherited review batches were caught by 6.2 Astra in 2/2 cases, versus Sol 1/2 and GPT-6 Astra 0/2; the always-BLOCKED control also catches 2/2. Latest v3.1 calibration has five overall winners, including 6.2 Astra and Sol, so no unique superiority follows. For bounded ideation, 6.2 Astra completed 0/3 calls within 480 seconds versus 3/3 for both older models; the requested Astra routing preference carries that limitation. Sources: November 17 `REPORT.md`, `ROLE_PROPOSAL.md`, `TASK_RESULTS.md`, and the committed `LEADERBOARD.md`. Remove all `google/gemini-*` entries from served roles on both Macs. Ricardo's reason: **“Spending OpenRouter money on Google is a waste”**. The code also excludes newly observed Google versions and prevents evidence or bootstrap from restoring them to served roles. Google routes, observations and dated scores remain readable as history. Each proposed private file removes 13 served entries across six affected roles; all eleven roles were inspected. Both current Macs have the same single `roles` map, with no `profiles` or `quality_roles`; the older model-pick docstring describing Mini/laptop profiles is stale. Applied 11-18 on both Macs: `~/.config/model-leaderboard/roles.json` replaced with the proposed file (backup `roles.json.bak-1118r`); `model-pick` now returns `gpt-6.2-astra` second in review, judge and ideation, and no role returns a `google/` model on either Mac. The routing code is on main, so the daily scorer, which writes `RH_ROLES_FILE` on the m5 and copies it to the M4, keeps Google out and seeds the 6.2 Astra slot when it regenerates the file. `~/.config/agent-backend` is unchanged on both Macs. Per-Mac diffs are in the `route62-1118` state archive. Did we find a new scientific result? No. This ships a routing fix and a limited cost comparison.