Loading the page.
Card · LLM for code · updated 2026-09-26
GPT-5.6
OpenAI · Sol / Terra / Luna family · previous generation, successor GPT-6 · effort up to max and pro mode · closed weights
72.7%
DeepSWE pass@1 · independent measurement
#10out of 27 models
Editorial review 2026-09-26
Editorial review 2026-09-26
OpenAI's previous generation in three tiers: Sol for complex code, Terra for everyday work, Luna for high-volume calls. In September the GPT-6 lineup replaced it. GPT-6 Sol costs half as much, $2/$10 vs $4/$20 per 1M tokens, and sits higher in the AA agents rating of Sep 25 (56.7 vs 54.6). On DeepSWE, GPT-5.6 Sol holds 72.7% and fourth place in the main table.
$4/$20
API Sol per 1M tokens · promo until at least Nov 21
1.05M
On all three tiers · output 128K
02.2026
Model knowledge cutoff
updated November 8, 2026
live · cron snapshot 24h · pass@1 and task price - deepswe.datacurve.ai · lines - Effort levels of a single model
Contents
1.0
Should you get it
Use for
- +a pipeline already tuned on GPT-5.6: it is not on the API shutdown list (Sep 26)
- +multi-file edits of the DeepSWE class: 72.3% in the AA run of Sep 25 vs 69.0% for GPT-6 Sol
- +a split across tiers: Sol for complex work, Terra for everyday work, Luna for high-volume work
Do not use for
- −new work: GPT-6 Sol is half the price per token and higher in the agents rating
- −a quality ceiling: in the AA agents rating, Claude Code with Opus 5.5 (66.0) and Codex with GPT-6 Astra (61.6) are ahead
- −self-hosting and open-weight scenarios: weights are closed
- −
In ChatGPT, Luna is available even on the free plan, Sol opens up with Plus, Sol Pro only on Pro. Via the API, Sol costs $4/$20 at the promo price until at least Nov 21, Terra $2/$12, Luna $0.20/$1.20.
2.0
Access and payment
Codex
codex /model gpt-5.6-sol
Codex CLI 0.144.0 or newer is required. GPT-6 Sol and Luna come first in the model list, GPT-5.6 is available during the transition. Cloud tasks on ChatGPT plans run on GPT-5.6 Sol.
OpenAI API
curl https://api.openai.com/v1/responses \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-d '{"model": "gpt-5.6-sol", "input": "..."}'The tier is picked by id: gpt-5.6-sol, -terra or -luna; the short gpt-5.6 points to Sol. Effort is set by the reasoning parameter, pro mode works only in the Responses API.
RU aggregator
curl https://api.proxyapi.ru/v1/chat/completions \
-H "Authorization: Bearer $PROXYAPI_KEY" \
-d '{"model": "openai/gpt-5.6-sol", "messages": [...]}'OpenRouter
curl https://openrouter.ai/api/v1/chat/completions \
-H "Authorization: Bearer $OPENROUTER_KEY" \
-d '{"model": "openai/gpt-5.6-sol", "messages": [...]}'One brain - different names · model ID by surface
| OpenAI API · Sol | gpt-5.6-sol | canonical, short alias gpt-5.6 |
| OpenAI API · Terra | gpt-5.6-terra | mid tier |
| OpenAI API · Luna | gpt-5.6-luna | fast tier |
| Codex | gpt-5.6-sol | picked via /model, CLI 0.144.0 or newer |
| ProxyAPI | openai/gpt-5.6-sol | |
| OpenRouter | openai/gpt-5.6-sol | unavailable to RU accounts |
For Sol in chat you need ChatGPT Plus or higher, Luna is available even on the free plan. For the API you need a platform.openai.com key with a balance.
3.0
Ranking position and benchmark variation
72.7%DeepSWE pass@1 · 09.2026Datacurve Sep 22, Sol max, fourth place in the main table. Above it are GPT-6 Astra, Claude Opus 5 and Gemini 3.8 Flash with 74%. GPT-6 Sol and Opus 5.5 have not been measured there yet
54.6AA agents rating · 09.2026version 1.5, Codex and Sol max, Sep 25: below GPT-6 Sol (56.7) and Astra (61.6), $6.35 per task
47AA Intelligence Index · 09.2026version 4.3, Sol max. GPT-6 Sol has 48, Claude Opus 5.5 has 58. AA flagged the model as outdated
64.6%SWE-bench Pro · 07.2026the vendor's claim (Sol) from the July release
Benchmarks differ - that is normal
Read it honestly: on DeepSWE, GPT-5.6 Sol is still above GPT-6 Sol (72.3% vs 69.0% in the AA run), while in the combined agents rating it loses because of Terminal-Bench 4.0 (37.4% vs 43.4%). The METR correction still applies. Sol has record reward hacking in the SWE eval, so its agentic percentages come with a caveat.
model ≠ agent
A model's score ≠ a tool's score: GPT-5.6 is measured both paired with Codex and with a simple open agent. Datacurve runs DeepSWE on Sol in max mode without Codex, the AA rating measures the Codex and Sol max pairing.
4.0
Family, versions, and modes
The 5.6 tiers shipped at the same time, on Jul 9. The next generation arrived in September: GPT-6 Astra on Sep 3, GPT-6 Sol and Luna on Sep 22. This card tracks the GPT-5.6 family, GPT-6 Astra has its own card.
GPT-6 AstraSep 3, 2026flagshipOpenAI's top model: $10/$50 via the API, 61.6 in the AA agents rating
GPT-6 Sol and LunaSep 22, 2026activethe successors to Sol and Luna: $2/$10 and $0.10/$0.50, half the price of GPT-5.6
GPT-5.6 SolJul 9, 2026active$4/$20 at the promo price until at least Nov 21. DeepSWE 72.7% (max), AA agents rating 54.6
GPT-5.6 TerraJul 9, 2026active$2/$12 after the price cut on Jul 30
GPT-5.6 LunaJul 9, 2026active$0.20/$1.20 after the Jul 30 cut, available in ChatGPT even on the free plan
GPT-5.504.2026activethe flagship before last, $5/$30: leaves ChatGPT and Codex on Oct 14, 2026, stays in the API
Reasoning modes
Spend is controlled by three levers: the effort ladder, standard or pro mode, and Ultra in Codex. The choice of lever changes the bill more than the choice between adjacent tiers.
01Effort: from none to maxsix steps on all three tiers, medium by default
02maxThe top step; DeepSWE (72.7%) and the AA agents rating were measured on it. As of the Jul 10 snapshot, before the price cut, max cost twice as much as xhigh: $8.39 vs $4.70 per task for +2 pp of quality.
03proA separate axis: the same price per token, but the model does more work and all of its tokens go on the bill. The bill grows noticeably.
04Ultra (Codex)Subagents split the task in parallel. In July OpenAI claimed 91.9% on Terminal-Bench 2.1 vs 88.8% solo. Not an API parameter.
5.0
Pricing and channels
Access channels · typical coding month, $
free
Free
Luna has no text message limit in ChatGPT chat, Sol does
$20
Plus
ChatGPT and Codex, GPT-5.6 Sol in Codex gets about 10–100 messages per 5 hours
$65
API
~10 DeepSWE-level tasks per month on Sol max ($6.46 per task)
$160
RU aggregator
API price
Sol $4 / $0.40 cached / $20 · Terra $2 / $0.20 / $12 · Luna $0.20 / $0.02 / $1.20 per 1M tokens
For Sol this is a promo price since Aug 21 instead of $5 / $30, and OpenAI is holding it at least through Nov 21, 2026. Terra got 20% cheaper and Luna 80% cheaper since Jul 30. The successors, out since Sep 22, cost half as much: GPT-6 Sol is $2 / $10, GPT-6 Luna $0.10 / $0.50. A cache write costs 1.25× the input, and a request longer than 272K tokens is twice as expensive on input and 1.5× on output.
OpenAI API docs · verified 2026-09-26
Where available
ChatGPT: Luna on every plan, including the free one; Sol from Plus up, Sol Pro only on Pro. API — all three tiers
On Free and Go, Luna runs with no limit on text chats; Terra is there in limited form in Work and Codex on the desktop, and in full from Plus up. In Codex the family stays on through the transition to GPT-6, and cloud tasks on ChatGPT plans run on GPT-5.6 Sol. OpenAI's estimate for Codex on Plus per 5 hours, not a hard limit: Sol 10–100 messages, Terra 25–200, Luna 250–2,000. GPT-6 Sol gets 15–150 there, but it is not in the regular ChatGPT chat yet.
ChatGPT plans · verified 2026-09-26
Modes
Effort none, low, medium (default), high, xhigh and max — on all three tiers; pro mode is a separate axis
Pro costs the same per token, but the model does more work and all of its tokens are billed. Ultra is a Codex product mode where subagents split the task in parallel; it is not an API parameter. In Codex you need CLI 0.144.0 or newer.
OpenAI reasoning guide · verified 2026-09-26
Prices go stale
Model price lists and aggregator exchange rates change more often than new versions are released. Each row above includes a verification date - do not trust figures without a date, including ours.
6.0
Data transparency and where your code goes
Vendor benchmarks
Sol: SWE-Bench Pro 64.6% · Terminal-Bench 2.1 88.8% · «DeepSWE v1.1» 72.7%
Vendor claim from the release. The independent DeepSWE snapshot on 10.07 matches the claim: 72.7% (Sol max). openai.com returns 403 - release figures were verified against two secondary reports.
MarkTechPost (release summary) · verified 2026-07-11
Independent snapshot
DeepSWE (Sep 22): Sol max 72.7% at $6.46 per task · AA agents rating v1.5 (Sep 25): Codex + Sol max 54.6
On DeepSWE only the rows at 74% beat Sol: GPT-6 Astra, Claude Opus 5 and Gemini 3.8 Flash. In the AA agents rating the pairing with GPT-6 Sol scored 56.7 at $2.99 per task, and with GPT-5.6 Sol 54.6 at $6.35. On DeepSWE in the same run the older model is ahead of the newer one (72.3% vs 69.0%), while on Terminal-Bench 4.0 it is behind (37.4% vs 43.4%). AA marked GPT-5.6 Sol as legacy. Its Intelligence Index v4.3 is 47, GPT-6 Sol's is 48. The July figures of 80 and 59 came from earlier versions of the indexes and do not compare with the current ones.
DeepSWE (Datacurve) · verified 2026-09-26
Benchmark caveat
METR: Sol has a record level of reward hacking in SWE evals among publicly tested models
The model more often than others "hacks" the evaluation criterion instead of solving the task honestly. This does not invalidate the figures, but agent percentages should be read with this caveat.
METR via Tech Times · verified 2026-07-11
Data
API traffic is not used for training by default; retention up to 30 days for abuse monitoring
ZDR - under an enterprise agreement. Cache ttl settings control cache lifetime, not retention.
OpenAI data guide · verified 2026-07-11
Weights
Closed weights - cannot be deployed locally
OpenAI's open lineup is separate (gpt-oss). Open-weight alternatives for code - Kimi K2.6, GLM-5.x; they lag behind flagships in coding.
OpenAI API docs · verified 2026-07-11
Track record · recent events
2026-09-22release
GPT-6 Sol and Luna are out — the successors to Sol and Luna
$2 / $10 and $0.10 / $0.50 per 1M tokens: half the price of GPT-5.6. In Codex, GPT-5.6 stays on through the transition
2026-09-03release
GPT-6 Astra — OpenAI's new top model
API $10 / $50 per 1M tokens, in ChatGPT from the Plus plan up
2026-08-21policy
Sol got cheaper: $4 / $20 instead of $5 / $30
a promo price, OpenAI is holding it at least through Nov 21, 2026
2026-07-30policy
Luna is 80% cheaper, Terra 20%
Luna now costs $0.20 / $1.20, Terra $2 / $12 per 1M tokens
2026-07-09release
GPT-5.6 family release: Sol, Terra, and Luna
GA immediately in ChatGPT, Codex, and API; vendor claim: SWE-Bench Pro 64.6%, Terminal-Bench 2.1 88.8% (Sol). Cache writes are now paid (1.25×)
2026-07-10release
DeepSWE snapshot: Sol max 72.7% - new #1
higher than Fable 5 xhigh (69.9%) at $8.39/task versus $13.41; Terra max 69.6%, Luna max 67.2%
If it does not work · top reasons
▸403 / unsupported_country - region block
▸model_not_found - wrong id
The short gpt-5.6 points to Sol; for Terra and Luna the tier is required: gpt-5.6-terra and gpt-5.6-luna. GET /v1/models returns the list of live models.
▸Codex does not show GPT-5.6
Update Codex CLI to 0.144.0 or newer. GPT-6 Sol and Luna come first in the model list, GPT-5.6 is picked via /model while OpenAI keeps it during the transition. Check the version: codex --version.
▸Bill increased without traffic growth
Check the cache and the mode: cache writes are billed (1.25× of input), and pro mode counts all the tokens of the work done. The same work on GPT-6 Sol costs half as much, $2/$10 vs $4/$20.
▸429 - subscription limit
Limits are counted in 5-hour windows, and ChatGPT Work and Codex share one allowance. Wait for the reset, buy extra credits or move bulk tasks to Luna. On Plus, OpenAI estimates its Codex allowance at 250–2,000 messages per 5 hours.
7.0
GPT-5.6 Sol or GPT-6 Sol: stay or switch
GPT-6 Sol took GPT-5.6 Sol's place in OpenAI's lineup two and a half months later. The fork is simple: the new model is half the price and slightly higher in the combined rating, the old one is still stronger on DeepSWE.
GPT-5.6 Sol
54.6
AA agents rating · Codex, max
GPT-6 Sol
56.7
AA agents rating · Codex, max
| Criterion | GPT-5.6 Sol | GPT-6 Sol |
|---|---|---|
| API price per 1M tokens | $4 / $20 · promo until at least Nov 21 | $2 / $10 |
| Task price in the AA agents rating | $6.35 | $2.99 |
| AA Intelligence Index | 47 · AA flagged it as outdated | 48 |
| Where available | API, ChatGPT with Plus, Codex during the transition | API, Codex and Work with Plus, not in the regular chat |
| Context | 1.05M | 1.05M |
- →Start new work on GPT-6 Sol: a task in the AA agents rating of Sep 25 costs half as much for a slightly higher score. Keep GPT-5.6 Sol where the pipeline is already tuned and multi-file edits of the DeepSWE class matter.
Other paths
GPT-5.6 or GPT-6 AstraOpenAI's new top model: 61.6 vs 54.6 in the AA agents rating, API $10/$50
GPT-5.6 or Claude Opus 5.5the same price per token ($4/$20), first place in the AA agents rating with 66.0
GPT-5.6 or GPT-5.5the flagship before last: pricier ($5/$30) and leaves ChatGPT and Codex on Oct 14
Frequently asked questions
▸Is there a model newer than GPT-5.6?
Yes, the GPT-6 line replaced it. GPT-6 Astra shipped Sep 3, GPT-6 Sol and Luna on Sep 22. GPT-6 Sol costs $2 / $10 per 1M tokens, half the price of GPT-5.6 Sol, and it stands higher in the AA agents rating of Sep 25: 56.7 against 54.6.
▸Should you move from GPT-5.6 to GPT-6?
For new work, yes: GPT-6 Sol is cheaper both per token and per task ($2.99 vs $6.35 in the AA agents rating of Sep 25). Staying on GPT-5.6 Sol makes sense if your pipeline is already tuned for it. It has no API shutdown date, and on DeepSWE in the same run it is ahead of the newer model (72.3% vs 69.0%).
▸How much does GPT-5.6 cost via API?
Sol costs $4 per 1M input tokens and $20 for output — that is a promo price, and OpenAI is holding it at least through Nov 21, 2026. Terra costs $2 / $12, Luna $0.20 / $1.20.
▸
▸Which is better for code, GPT-5.6 or Claude Opus 5.5?
Opus 5.5 is stronger for code: in the AA agents rating of Sep 25 the pairing of Claude Code + Opus 5.5 scored 66.0 against 54.6 for Codex + GPT-5.6 Sol. Their per-token price is the same, $4 / $20 per 1M. But a task on GPT-5.6 Sol costs half as much, $6.35 vs $13.04.
▸Which tier to pick: Sol, Terra or Luna?
Sol is taken for agentic coding and hard tasks, Terra for everyday work, Luna for cheap high-volume calls. Context and knowledge cutoff are the same across all three; the difference is price and depth. GPT-6 has no Terra tier, and GPT-6 Sol costs less than it does: $2 / $10 vs $2 / $12.
▸What are Ultra and Programmatic Tool Calling?
Ultra is a Codex mode where subagents split the task in parallel. In July OpenAI claimed 91.9% for Ultra on Terminal-Bench 2.1 against 88.8% for Sol on its own. Programmatic Tool Calling lets the model drive tools from JavaScript code in an isolated environment with no network, and OpenAI claims token savings of 38–63.5% among early customers.
▸Can it be deployed locally?
No, the weights are closed; OpenAI's open lineup is the separate gpt-oss models. If you need self-hosting - look at the open-weight class: Kimi K2.6, GLM-5.x, DeepSeek V4. They lag behind the flagships on code, but they run on your own hardware with no restrictions.
▸Does the model train on my code?
For the API - no, not by default: OpenAI's official policy allows retention for up to 30 days for abuse monitoring. In ChatGPT personal plans, training is controlled through the Data Controls setting.
More on the topic
articleGPT-6 Sol and Luna are 50% cheaper than GPT-5.6the successors to Sol and Luna: pricing and accessarticleGPT-5.6 Sol got cheaper for three monthswhere the $4/$20 price comes from and how long it lastsguideLLM ranking for codingthe live index: where GPT-5.6 sits among all modelsguideTelegram@shilovtechlive author notes: releases, limits, tricks - unfiltered
Entity · gpt-5-6
Facts · 14, each with a source and date
Card editorial review · 2026-09-26
Facts verified · 2026-09-26