Repository navigation
GPT-6.1 Terra: follow-up data hunt until AA/provider data is sufficient #1249
Description
Context
GPT-6.1 Terra was released today (2026-10-26). A first data sweep is underway tonight (AA page if it exists, OpenAI's launch claims, models.dev, provider catalogs, third-party reports) and will land as a dated note in ~/workspace/ai-subscriptions/. This issue owns the FOLLOW-UP trail: release-day data for a new model is always incomplete, so keep hunting on a cadence until the picture is sufficient or it's clear no more data is coming.
Claims to verify (attributed to OpenAI, per Drew's recollection — treat as unverified until sourced)
- Terra costs half what the previous Terra cost.
- Terra is cheaper than GPT-6.1 Sol (Drew said "Soul" — almost certainly Sol).
Checklist
- Tonight's sweep reviewed: what exists, what's missing (esp. Artificial Analysis measurements: II / TTFT / speed / cost per task per effort rung).
- Recheck 2026-10-27 (tomorrow): AA Terra page, models.dev entry, provider catalogs (OpenCode Go, OpenAI API).
- Recheck ~2026-11-02 (next week): AA measurements, any third-party benchmarks/anecdotes.
- Weekly thereafter until: AA data exists AND at least one provider door we hold lists Terra — then evaluate Terra for pool candidacy (which family slot does it take: Luna's? a Terra-plan pool?) and close.
- Close early only if: OpenAI/providers show Terra was withdrawn or renamed, or 6 weeks pass with no AA measurement (record that as the finding).
Storage
Findings go in as dated notes (tonight: ~/workspace/ai-subscriptions/terra-2026-10-26.md; later: dated follow-ups alongside it), and once the append-only model-data pipeline (companion issue) exists, as measurement events + notes there. Record absences too — "AA still has no Terra page as of " is a finding.
Acceptance criteria
- Each scheduled recheck (2026-10-27, ~2026-11-02, then weekly) lands as a dated note in
~/workspace/ai-subscriptions/— including recorded absences (e.g. 'AA still has no Terra page as of '). - The two attributed OpenAI claims (Terra at half the previous Terra price; cheaper than GPT-6.1 Sol) are each marked verified-with-source or unverified at every recheck — never repeated as fact without a source.
- The issue closes only when AA data exists AND at least one provider door we hold lists Terra (with the pool-candidacy evaluation recorded), or via the stated early-close conditions (withdrawn/renamed, or 6 weeks with no AA measurement recorded as the finding).
Activity
developing-today commentedon Oct 27, 2026
CODE-374
dezren39 commentedon Oct 27, 2026
Tonight's sweep (2026-10-26) — release-day findings
Note:
~/workspace/ai-subscriptions/terra-2026-10-26.mdStill open (the follow-up trail continues)
Cadence unchanged: recheck 2026-10-27, ~2026-11-02, then weekly per the issue body.
dezren39 commentedon Oct 27, 2026
Intake review
Related: #1248 (the append-only model-data pipeline — once it exists, findings land there as measurement events + notes as well as the dated files).
Notes: A cadence-tracking issue, not a one-shot fix: release-day model data is incomplete by nature. The checklist in the body is the work plan; this pass only normalized labels and formalized the closing conditions as acceptance criteria.
Risks: The two price claims are Drew's recollection of OpenAI's launch material and are explicitly unverified — keep them attributed until a first-party source is captured. Cadence items like this go stale silently; the weekly recheck needs an owner (a scheduled sweep slot) or it will drift.
Suggested approach: Attach the rechecks to an existing scheduled research/token-review cadence rather than a new job; each recheck appends one dated note and ticks the checklist.
dezren39 commentedon Oct 27, 2026
Per-rung latency — pulled from AA 2026-10-27 ✅
The missing AA latency numbers (Gap §5.1 in the research note) are now in. TTFT and output speed were read directly from each rung's own Artificial Analysis model page (each page's FAQ, "based on OpenAI's API"); output tokens/task are the AA figures already in the note. Decode time = tokens ÷ speed (AA's own "Time per Intelligence Index Task" definition); per-task total = TTFT + decode (computed, not AA-published).
Routing-critical number: medium TTFT is 1.43s — essentially tied with low (1.28s) and non-reasoning (1.27s); its per-task total (~65.6s) is decode-bound, not latency-bound. Nothing in medium's latency argues against the Terra #medium seat (workhorse + specialist, weight 1) taken on 2026-10-26 pending exactly this number.
Other findings:
Checklist impact: the AA-measurements part of the first checklist item (II / TTFT / speed / cost per task per effort rung) is now complete — II and cost/task landed 2026-10-26, TTFT and speed landed with this pull. Full table + per-rung source URLs appended to
~/workspace/ai-subscriptions/terra-2026-10-26.mdunder "Per-rung latency (pulled 2026-10-27)". Remaining open items on this issue are unchanged (models.dev verification, provider-door listings, pool-candidacy evaluation on the recheck cadence).dezren39 commentedon Oct 27, 2026
Recheck 2026-10-27 (due recheck) — four items, all executed
1. models.dev direct check — VERIFIED, matches records exactly
Fetched https://models.dev/api.json directly (5,956,211 bytes). 16 providers list a
gpt-6.1-terraentry: openai, azure, azure-cognitive-services, commandcode, openrouter, github-copilot, edenai, kilo, llmgateway, llmgateway-providers, lowrouter, merge-gateway, nano-gpt, aimlapi, requesty, vercel.The
openaientry matches everything recorded on release day, field for field: context 1,050,000, input cap 922,000, output 128,000; cost $1 input / $5 output / $0.05 cache_read / $1.25 cache_write per 1M; >272,000-context tier $2 / $7.50 / $0.10; release 2026-10-26, knowledge 2026-06-01, familygpt-terra. No discrepancy on any core field.Two variations worth recording, neither a discrepancy for OpenAI's own spec: (a) the effort-rung lists differ per provider (openai/azure: none/low/medium/high/xhigh/max; commandcode: low–max, no "none"; requesty: no xhigh; aimlapi: low/medium/high only) — upstream catalogs disagree on which rungs each provider exposes; (b) two "Terra Pro" entries (openrouter, kilo — pro reasoning mode, same underlying model, same $1/$5) and one vercel
gpt-6.1-terra-fast($2/$10 priority tier). Curiosity: models.dev lists zero providers forgpt-6.1-luna— that model id appears not to exist in the catalog at all.2. OpenCode Go listing — NO (two independent sources agree)
opencode-goprovider: 37 models; the only GPT entries aregpt-5.6-lunaandgpt-6-luna. No GPT-6.1 model of any kind.Implication: Terra cannot ride the Go quota. Our opencode-go leaves (in the luna/sol plan pools) are Luna-only. Separately, the PAYG
opencode(Zen) catalog listsgpt-6.1-solbut not Terra — plan-pool irrelevant, noted for completeness.3. AA snapshot freshness — unchanged except one revision
Max page (artificialanalysis.ai/models/gpt-6-1-terra) re-read today. Unchanged: Intelligence Index 48 (v4.3.2, same scale), output speed 90.9 t/s, pricing $1 / $5 (blended $0.73), context 1.0M, 130M output tokens, 6 providers, release 2026-10-26. No new rungs, no new benchmark data on the page.
Changed — max TTFT revised upward: the page's FAQ now reads: "GPT-6.1 Terra (Max) has a time to first token (TTFT) of 133.64s (based on OpenAI's API)". The launch-day and 2026-10-27 pulls both read 77.96s; a text search of the current page confirms 133.64 present and 77.96 absent. So AA has re-measured/revised max TTFT ~1.7×. Effect on the note's per-rung table: max per-task total becomes ~519.7s (133.64 + 386.1 decode) instead of 464.1s — it strengthens the "max is escalation-only" conclusion and touches nothing about medium (the routing-critical rung). Only the max page was rechecked today; the other five rungs' TTFTs were not rechecked — the ~11-02 pass should re-pull all six.
4. Live-door availability spot-check — NO (neither proxy advertises it)
Read-only
/v1/modelsGETs on the laptop proxies (2026-10-27):gpt-5.6-terrafamily; 6.1 ids = only thegpt-6.1-solfamily. Nogpt-6.1-terra.gpt-5.6-terra; 6.1 = onlygpt-6.1-sol. Nogpt-6.1-terra.Notable asymmetry: models.dev's commandcode catalog does list Terra, but the live commandcode proxy's list does not — proxy lists are curated subsets, consistent with prior findings. (The Codex CLI itself still serves Terra by direct invocation, probe-verified 2026-10-26; only the proxy advertisement is absent.) Nothing was restarted or changed.
Closing assessment
The acceptance conditions are nearly met, but I am not closing this issue yet:
terra-planpool with workhorse and specialist#mediumseats (config PRs feat(opencode): set up GPT-6.1 Terra (probe-verified via Codex) #1296/feat(opencode): Terra workhorse weight 1 → 2 (medium TTFT 1.43s) #1312/feat(opencode): pause claude-code (free tier) + terra-plan door correction #1316), justified by medium's 1.43s TTFT and frontier position. Recording that here as the evaluation.--jsontranscript output in world line mode #563, merged) lists Terra, andterra-planaddresses it explicitly — but no live proxy we run advertises it yet, and Go carries no 6.1 at all.Recommendation: cadence stands. One more recheck ~2026-11-02, focused on: (a) re-pull all six AA rung TTFTs (AA is revising latency figures), (b) re-check both live proxy lists, (c) re-check Go. Close then if a held live door lists it, or if the fork-catalog + explicit terra-plan entries are accepted as satisfying the "door we hold" condition. Early-close conditions (withdrawn/renamed, 6 weeks without AA measurement) are not triggered — the model is healthy and expanding in catalogs.
1 remaining item
dezren39 commentedon Oct 28, 2026
🤖 Agent note · parked
~/workspace/ai-subscriptions/Cadence state, verified from this issue's own record (2026-10-28):
terra-2026-10-26.md) — checklist item 1 (tonight's sweep reviewed, incl. AA gaps) is substantively done; per-rung TTFT/speed gaps were then pulled from AA on 2026-10-27.gpt-6.1-terra, including openai. So one close ingredient (a provider door we hold listing Terra) already looks satisfiable — but the issue does not close yet: the pool-candidacy evaluation has not been recorded, and the cadence's next steps come first.Route: task 2515 parked until 2026-11-02 (the next recheck window) rather than completed-and-requeued daily, which would just churn triage against a future date. Issue labels:
queue/parked+status/parked,work/assignedstays with the parked task. On resume, run the Nov-2 recheck and record it here + as a dated note.dezren39 commentedon Nov 2, 2026
Recheck 2026-11-02 (due recheck) — closing recheck: conditions met
Full dated note:
~/workspace/ai-subscriptions/terra-2026-11-02.md1. AA — Index v5.0 is live; all six rungs re-pulled (the (a) item)
AA replaced Index v4.3.2 with v5.0 — a new scale, not comparable to the v4.3.2 figures in the release-day note. Per-rung, read from each rung's own AA page today (price $1/$5 on every rung, context 1.0M):
#mediumspecialist seat in the 2026-10-27 candidacy evaluation — now reads 4.90s. Filed as follow-up flag #4390 for the pool owners (#1673 re-baselined II and $/task to v5 but not this TTFT premise). It does not block closing here.2. Live doors (the (b) item) — BOTH proxies now list Terra ✅
Read-only
/v1/modelson the laptop today, each verified with a second independent read:gpt-6.1-terrafamily — base,-low,-medium,-high,-xhigh,-max,-ultra.gpt-6.1-terra.On 2026-10-27 neither advertised Terra at all.
3. OpenCode Go (the (c) item) — still NO
No
gpt-6.1-terrakey in theopencode-goblock of today's models.devapi.json. Terra still cannot ride the Go quota; the doors that opened are the proxies.4. models.dev + claims
openaientry forgpt-6.1-terrais field-for-field identical to the 2026-10-27 record ($1/$5/$0.05/$1.25, ctx 1,050,000, >272K tier $2/$7.50/$0.10). Exact-key entries enumerated in at least nine provider catalogs today.Closing assessment
The acceptance close condition is met, on the strong form the 2026-10-27 assessment asked for: AA data exists (refreshed in full today) AND held live doors list Terra (both proxies, verified twice) AND the pool-candidacy evaluation is on record (Terra took its own
terra-planpool; seats per config PRs #1296/#1312/#1316 — with the #4390 TTFT-premise flag handed off). Early-close conditions were not triggered. Closing #1249; the cadence ends. Future Terra data movement belongs to #1248 and the pool registry.dezren39 commentedon Nov 2, 2026
🤖 Agent note · done