Skip to content

KB — GPT-5.6 (Sol / Terra / Luna)

Research note

This page is background research kept for the project's own reference. It is not user documentation, it is not kept up to date with every release, and model or product details in it may be out of date. For how anti-hall works today, start at the home page.

OpenAI's GPT-5.6 model family, its three named tiers, pricing, API/Codex CLI availability, and this machine's local Codex CLI gap. Compiled 2026-07-11 from a deep-research pass (19 sources fetched, 25 claims adversarially cross-verified, 23 confirmed / 2 refuted / 0 unverified) plus a primary-source re-verification pass against OpenAI's own developer docs, and a local check of the installed Codex CLI on this machine.

Confidence & provenance

Three confidence bands are used throughout this doc — read them literally, they are load-bearing:

  • PRIMARY-CONFIRMED — fetched HTTP 200 directly from an OpenAI-controlled docs domain (developers.openai.com/*, which redirects 308 to learn.chatgpt.com/docs/*) during the re-verification pass. Stated as fact.
  • SECONDARY-CONFIRMED — corroborated by multiple independent secondary sources (GitHub's own changelog, Wikipedia, tech press) but the primary OpenAI page for this specific claim was not successfully machine-read. Labeled explicitly where used.
  • PRESS/BENCHMARK — third-party benchmark aggregator or press reporting, not an OpenAI-published number. Labeled explicitly, never asserted as OpenAI-official.

The 403 caveat: OpenAI's own announcement pages — openai.com/index/gpt-5-6/, openai.com/index/previewing-gpt-5-6-sol/, and the help.openai.com preview article — returned HTTP 403 to automated fetch (bot protection, not 404 — the pages are live and indexed; content was corroborated via search-result extraction and independent secondary sources, not read verbatim by a fetcher). Anything resting solely on those pages is marked secondary-confirmed below, even though OpenAI is the ultimate origin.

1. The three tiers

GPT-5.6 is a generation, not a single model; Sol / Terra / Luna are three named, independently-priced tiers within it. OpenAI's own framing (from the developers.openai.com docs, PRIMARY-CONFIRMED): "the number identifies a model's generation, while Sol, Terra, and Luna identify durable capability tiers that can advance on their own cadence."

Tier Official API model ID Positioning Reasoning-effort notes
Sol gpt-5.6-sol Flagship/frontier — "strongest capability for complex coding, computer use, research, and cybersecurity." Bare gpt-5.6 is an alias that routes to Sol: "The gpt-5.6 alias routes requests to GPT-5.6 Sol." Only tier documented with a max reasoning-effort level and an "ultra" mode (press-reported as parallel-subagent execution — see §7 caveat).
Terra gpt-5.6-terra Balanced everyday/agentic-coding tier — "competitive with GPT-5.5 at lower cost." Standard reasoning-effort range (not documented as offering max).
Luna gpt-5.6-luna Fastest, lowest-cost tier in the family. Standard reasoning-effort range; targets latency-sensitive workloads.

Source for the tier table: developers.openai.com/api/docs/models/gpt-5.6-sol and developers.openai.com/api/docs/models — PRIMARY-CONFIRMED.

2. Pricing

Per 1M tokens, standard rates (developers.openai.com/api/docs/pricing — PRIMARY-CONFIRMED):

Tier Input Output Cached input
Sol $5.00 $30.00 $0.50
Terra $2.50 $15.00 — (not separately re-verified; treat as proportionally discounted pending direct check)
Luna $1.00 $6.00 — (same caveat)

Additional billing rules (PRIMARY-CONFIRMED, from the same pricing page and the Sol model-card page):

  • Cache writes are billed at 1.25× the uncached input rate.
  • Long-context surcharge: prompts whose input exceeds 272K tokens are billed at 2× input / 1.5× output for the entire request (not just the excess).

3. API access

Per developers.openai.com/api/docs/models/gpt-5.6-sol and developers.openai.com/api/docs/models (PRIMARY-CONFIRMED):

  • Endpoints: Chat Completions, Responses, Realtime, Batch.
  • Features: streaming, function calling, structured outputs.
  • Alias behavior: the bare gpt-5.6 string is not its own model — it is an alias that resolves to gpt-5.6-sol. Callers who want Terra or Luna must name the tier explicitly (gpt-5.6-terra / gpt-5.6-luna).
  • Context-window size and per-tier rate limits were not confirmed from a primary source during re-verification (an earlier research pass extracted a 1,050,000-token context / 128,000-token max-output figure for Sol from a search snippet, but this was not independently re-fetched from the model card during the primary pass) — not stated as fact here; verify directly against developers.openai.com/api/docs/models/gpt-5.6-sol before relying on it.

4. Availability & rollout

  • GA date — SECONDARY-CONFIRMED, not primary-read: OpenAI's own two announcement pages (openai.com/index/gpt-5-6/, openai.com/index/previewing-gpt-5-6-sol/) 403'd to direct fetch. The 2026-07-09 general-availability date (across ChatGPT, Codex, and the OpenAI API, following a limited preview beginning ~2026-06-26) is corroborated by:
  • GitHub's own Copilot changelog (PRIMARY-CONFIRMED for the Copilot-rollout claim specifically — see next bullet): "Release Date: July 9, 2026."
  • Wikipedia's GPT-5.6 article (secondary).
  • Multiple press outlets — TechCrunch, CNBC, MarkTechPost, Vellum (secondary, independently converging on the same date).
  • GitHub Copilot rollout — PRIMARY-CONFIRMED: github.blog's own changelog entry (dated 2026-07-09) states GPT-5.6 Sol, Terra, and Luna became available in GitHub Copilot starting that date. This is GitHub's own official changelog, so it is treated as primary for the Copilot-specific claim (independent of the 403'd OpenAI pages).
  • Codex CLI selectability — PRIMARY-CONFIRMED: learn.chatgpt.com/docs/models (the page developers.openai.com/codex/models 308-redirects to — an OpenAI-owned docs domain, fetched HTTP 200) documents Codex CLI model selection for GPT-5.6 verbatim, including: 
    codex -m gpt-5.6-sol
codex --model gpt-5.6
    alongside gpt-5.5, gpt-5.4, gpt-5.4-mini, and gpt-5.3-codex-spark as other selectable models. /model is documented as the interactive-session switch.

5. Local Codex availability caveat (this machine)

Verified on this machine 2026-07-10, not a general OpenAI fact:

  • The installed codex CLI is v0.143.0, default model gpt-5.5.
  • Its local cache file ~/.codex/models_cache.json (fetched 2026-07-10) lists only gpt-5.5, gpt-5.4, gpt-5.4-mini, and gpt-5.3-codex-spark — GPT-5.6 (any tier) is absent from the interactive /model picker on this install.
  • This matches a filed upstream bug, github.com/openai/codex/issues/31873 (filed against v0.143.0): "/model does not list GPT-5.6 models that are available via -m" — i.e., GPT-5.6 is invocable via codex -m gpt-5.6-sol (per §4 above), it is just missing from the picker/cache on this version.
  • Practical consequence: invoking an uncached model id on this install emits "Model metadata not found. Defaulting to fallback metadata" — reasoning-effort and tool defaults may be degraded relative to a fully-cached model until the picker/cache bug is fixed upstream.

6. Coding/agentic benchmarks (third-party, not OpenAI-official)

Caveat first: OpenAI's own benchmark/announcement pages were not readable (403, see §0 confidence note above). Every number below comes from third-party benchmark aggregators or press, reported as such, not asserted as OpenAI-official:

  • Artificial Analysis Coding Agent Index v1.1 (per MarkTechPost's reported table and independently per artificialanalysis.ai's own article): Sol scores 80, ahead of the reported field — press coverage frames this as Sol leading the index.
  • SWE-Bench Pro (press-reported, via vellum.ai / Artificial Analysis coverage): Sol 64.6%, versus a reported 80.3% for a Claude codename used by these outlets ("Claude Mythos 5") — a roughly 15–16 point gap in Sol's favor for the competitor. Vellum's own reporting notes OpenAI did not itself publish Sol's SWE-Bench Pro figure, and the Claude codenames used by these aggregators are the outlets' own naming, not confirmed Anthropic product names.
  • Treat both figures as third-party benchmark claims, not first-party OpenAI results, and re-verify against a transparent/reproducible harness before citing them as settled.

7. Relevance to anti-hall's Codex port

Brief pointer only — this doc does not design a migration. Anti-hall's current Codex port (plugins/anti-hall/codex/) routes debate/validation/planning seats to gpt-5.5, implementation to gpt-5.4, and mechanical/cheap work to gpt-5.4-mini (default), per plugins/anti-hall/codex/skills/*/SKILL.md and plugins/anti-hall/codex/README.md. GPT-5.6 Sol/Terra/Luna are the natural migration targets for those same three seats respectively once the local picker/cache gap (§5) is resolved upstream and the model is broadly available on this install — but the actual migration plan is out of scope for this KB entry and belongs in a separate design doc.

8. Sources

Official/primary sources are marked [OFFICIAL]. OpenAI's own announcement pages that 403'd to direct fetch (content corroborated via search/secondary sources, not read verbatim) are marked accordingly.

  1. developers.openai.com/api/docs/models/gpt-5.6-sol — [OFFICIAL] Sol model card (tier ID, positioning, endpoints/features).
  2. developers.openai.com/api/docs/models — [OFFICIAL] full model list incl. all three GPT-5.6 tier IDs.
  3. developers.openai.com/api/docs/pricing — [OFFICIAL] per-tier pricing, cache-write multiplier, >272K surcharge.
  4. developers.openai.com/codex/models (308-redirects to learn.chatgpt.com/docs/models) — [OFFICIAL] Codex CLI model-selection syntax (-m, --model, /model) and per-tier availability matrix.
  5. learn.chatgpt.com/docs/models — [OFFICIAL] the redirect target of #4; fetched HTTP 200; documents codex -m gpt-5.6-sol / codex --model gpt-5.6 alongside gpt-5.5/5.4/5.4-mini/5.3-codex-spark.
  6. developers.openai.com/codex/changelog (308-redirects to learn.chatgpt.com/docs/changelog) — [OFFICIAL] Codex CLI version history; v0.143.0 (2026-07-08) Bedrock GPT-5.6 support entry, v0.144.0 (2026-07-09) naming-display fix entry.
  7. github.blog/changelog/2026-07-09-openais-gpt-5-6-sol-terra-and-luna-are-now-available-in-github-copilot/ — [OFFICIAL] GitHub's own changelog confirming the Copilot rollout date and the three tier names.
  8. openai.com/index/gpt-5-6/ — [official, not machine-readable — 403] OpenAI's primary GPT-5.6 announcement page; content triangulated via search extraction and secondary sources only.
  9. openai.com/index/previewing-gpt-5-6-sol/ — [official, not machine-readable — 403] OpenAI's limited-preview announcement page; same caveat as above.
  10. help.openai.com/en/articles/20001325-a-preview-of-gpt-56-sol-terra-and-luna — [official, not machine-readable — 403] OpenAI support-center preview article; direct fetch 403'd, content corroborated via search extraction.
  11. github.com/openai/codex/releases/tag/rust-v0.143.0 — OpenAI's own Codex CLI GitHub release page, corroborating the v0.143.0/2026-07-08 date.
  12. github.com/openai/codex/issues/31873 — filed upstream bug (this machine's install, v0.143.0): /model omits GPT-5.6 from the picker despite -m support.
  13. github.com/openai/codex/issues/31905 — community-filed Codex issue referencing GPT-5.6/Bedrock rollout, used as corroborating (forum-quality) evidence.
  14. github.com/openai/codex/issues/31869 — community-filed Codex issue referencing GPT-5.6 availability, corroborating (forum-quality) evidence.
  15. en.wikipedia.org/wiki/GPT-5.6 — secondary encyclopedia summary corroborating GA date and tier structure.
  16. techcrunch.com/2026/07/09/openai-launches-its-new-family-of-models-with-gpt-5-6/ — press coverage of the 2026-07-09 GA rollout.
  17. www.marktechpost.com/2026/07/09/openai-releases-gpt-5-6-a-three-tier-model-family-with-programmatic-tool-calling/ — press coverage; source of the reported Coding Agent Index and SWE-Bench Pro table.
  18. www.vellum.ai/blog/gpt-5-6-benchmarks-explained — third-party benchmark analysis blog; explicitly notes OpenAI did not itself publish the SWE-Bench Pro number for Sol.
  19. artificialanalysis.ai/articles/gpt-5-6-has-landed — third-party benchmark aggregator's own published article (independent confirmation of the Coding Agent Index scores).
  20. www.datacamp.com/blog/claude-sonnet-5-vs-gpt-5-6 — press/blog head-to-head comparison, used only for cross-checking benchmark figures, not as a primary claim source.
  21. venturebeat.com/technology/openai-unveils-gpt-5-6-sol-terra-and-luna-models-but-only-accessible-to-limited-preview-partners-for-now-per-us-gov — press coverage of the limited-preview phase preceding GA.
  22. www.axios.com/2026/06/26/openai-gpt-sol-terra-luna-trump — press coverage of the ~2026-06-26 limited-preview start date.
  23. lmcouncil.ai/benchmarks — third-party benchmark listing site; cited here specifically because a claim sourced from it (that GPT-5.6 was absent from its comparison page as of 2026-07-01) was tested and refuted (0-3 vote) during adversarial verification — kept in the source list for that provenance, not as a claim source.
  24. www.morphllm.com/best-ai-model-for-coding — blog-quality source consulted during the coding-benchmark research angle; not used for a load-bearing claim in this doc.
  25. codex.danielvaughan.com/2026/06/03/gpt-5-6-codex-cli-canary-signals-developer-readiness-guide/ — independent blog tracking pre-GA Codex CLI canary signals for GPT-5.6.
  26. techsy.io/en/blog/gpt-5-6-leak — blog covering pre-announcement leak chatter; consulted specifically for the "leak vs. official" skeptical angle, superseded by the confirmed official sources above.

9. Facts dropped or down-labeled for lack of verification

  • Context window / max output size (a 1,050,000-token context / 128,000-token max-output figure for Sol) — dropped from the fact sections above. It surfaced in one search-snippet extraction during the initial research pass but was flagged by that same research as not confirmed from a primary source (explicit open question in the research output), and it was not independently re-verified during the primary-source re-check pass. Do not treat it as established; re-fetch developers.openai.com/api/docs/models/gpt-5.6-sol directly before relying on it.
  • Per-tier rate limits — not surfaced by any source in this research pass; omitted entirely rather than guessed.
  • Terra/Luna cached-input pricing — only Sol's cached-input rate ($0.50) was directly re-verified against the primary pricing page during the re-verification pass; Terra/Luna cached rates are left uncited above rather than extrapolated.
  • The GA date itself — down-labeled to SECONDARY-CONFIRMED throughout this doc (not stated as OpenAI-primary) because the two OpenAI announcement pages that would be the primary source for it both returned HTTP 403 to direct fetch during this research; only secondary/press corroboration was available.
  • "Sol-exclusive max/ultra mode" as a hard architectural fact — the underlying adversarial verification pass recorded a 2-1 (not unanimous) vote on this specific detail. Kept in §1's table because it is the majority-confirmed reading, but flagged here as the one item in this doc with a dissenting vote rather than unanimous confirmation.
  • Regional/enterprise gating during the June 26 limited preview (reported by press as restricted per a US-government request) — not included in the fact sections above; this research pass did not verify it against a primary source and it is unrelated to the GA-era facts this doc otherwise states.