Loading the page.
Card · LLM for code · updated 2026-07-11
GPT-5.6
OpenAI · Sol / Terra / Luna family · #1 AA Coding Agent · reasoning up to xhigh and pro · closed weights
72.7%
DeepSWE pass@1 · independent measurement
#2out of 16 models
Editorial review 2026-07-11
Editorial review 2026-07-11
GPT-5.5's three-tier successor family: Sol - the flagship for agentic coding, Terra - the balance for everyday work, Luna - cheap high-volume calls. Sol is the new #1 on our anchor DeepSWE: pass@1 72.7% (max, 10.07 snapshot) versus 69.9% for Fable 5 at $8.39/task versus $13.41. AA Coding Agent Index also puts it first (80, #1).
$5/$30
api per 1M tokens · sol; terra $2.50/$15, luna $1/$6
1.05M
across all three tiers · output 128K
02.2026
model knowledge cutoff
Yes
Direct payment from Russia
from Russia: officially unavailable · ruble aggregator; OpenRouter unavailable
updated September 22, 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
- +primary brain for agentic coding: Codex, autonomous pipelines, long tasks
- +frontier economics: $8.39/DeepSWE task versus $13.41 for Fable 5 (10.07 snapshot)
- +tier routing: Sol - complex, Terra - everyday, Luna - high-volume
- +bulk work against a clear spec: migrations, data processing, repetitive edits
Do not use for
- −quality ceiling for the hardest tasks: Fable 5 ranks higher in AA Intelligence (60 versus 59)
- −self-hosting and open-weight scenarios: weights are closed
- −personal-data processing under Federal Law No. 152-FZ: servers are overseas
- −direct payment from Russia - does not work; OpenRouter is also closed for RU accounts
ChatGPT subscription: Sol from the Plus plan (Plus - effort medium/high only); Free and Go do not get GPT-5.6 in chat. Via API, all three tiers: Sol $5/$30, Terra $2.50/$15, Luna $1/$6. From Russia - RU aggregators in rubles.
2.0
Access and payment
Codex
codex /model gpt-5.6-sol
Requires Codex CLI 0.144.0 or later; all three tiers have effort controls, Ultra mode - from the Plus plan.
OpenAI API
curl https://api.openai.com/v1/responses \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-d '{"model": "gpt-5.6-sol", "input": "..."}'Tier is selected by id: gpt-5.6-sol / -terra / -luna; effort and pro mode - via the reasoning parameter.
RU aggregator
curl https://api.proxyapi.ru/openai/v1/chat/completions \
-H "Authorization: Bearer $PROXYAPI_KEY" \
-d '{"model": "gpt-5.6-sol", "messages": [...]}'OpenAI-compatible format: only base_url and key change. Rubles, no foreign card, ~1.5-2× markup.
OpenRouter
curl https://openrouter.ai/api/v1/chat/completions \
-H "Authorization: Bearer $OPENROUTER_KEY" \
-d '{"model": "openai/gpt-5.6-sol", "messages": [...]}'Blocked for RU accounts since ~05.2026: OpenAI models are on the Restricted list. A route for foreign accounts.
One brain - different names · model ID by surface
| OpenAI API · Sol | gpt-5.6-sol | canonical; no dated snapshots |
| OpenAI API · Terra | gpt-5.6-terra | mid tier |
| OpenAI API · Luna | gpt-5.6-luna | fast tier |
| Codex | gpt-5.6-sol | selection via /model; CLI from 0.144.0 |
| OpenRouter | openai/gpt-5.6-sol | unavailable to RU accounts |
For a subscription - a ChatGPT account with a plan from Plus; for API - a platform.openai.com key with balance. From Russia, neither can be paid for directly.
Availability from Russia
OpenAI blocks Russia by IP and cards. OpenRouter also blocked OpenAI models for RU accounts from ~05.2026 - the old recipe no longer works. Working routes: RU aggregators in rubles without a foreign card - ProxyAPI, AITUNNEL, VseGPT (~1.5-2× markup) - or an account and card from another country.
3.0
Ranking position and benchmark variation
72.7%DeepSWE pass@1 · 07.2026independent measurement deepswe.datacurve.ai, Sol max - #1; Fable 5 xhigh - 69.9%
80AA Coding Agent Index · 07.2026independent measurement, #1: ahead of Fable 5 and Opus 4.8; $1.04/task (Sol)
64.6%SWE-bench Pro · 07.2026vendor claim (Sol); Anthropic's claim for Fable 5 is higher - 80.3%
Benchmarks differ - that is normal
Read honestly: both independent measurements favor Sol - DeepSWE 72.7% versus Fable 5's 69.9% and AA Coding Agent 80, but on AA Intelligence Fable is higher (60 vs 59), while on SWE-bench Pro Anthropic's claim beats OpenAI's claim (80.3% vs 64.6%). The caveat remains: METR recorded record reward hacking by Sol in SWE evals - read agent percentages with this qualification.
model ≠ agent
Model score ≠ tool score: GPT-5.6 is measured both paired with Codex and solo. DeepSWE fixes the configuration: 72.7% is Sol in max mode; xhigh delivers 70.7% at half the cost ($4.70 vs $8.39 per task).
4.0
Family, versions, and modes
The 5.6 tiers release simultaneously, not sequentially: the family timeline reads as a price ladder. The card tracks the GPT-5.6 family; previous generations move to the archive as they are deprecated.
GPT-5.6 Sol09.07.2026flagshipflagship $5/$30: #1 DeepSWE (72.7% max) and AA Coding Agent (80); modes up to max and pro
GPT-5.6 Terra09.07.2026activebalanced $2.50/$15: DeepSWE 69.6% (max) - at Fable 5 level for ~⅓ of the task price
GPT-5.6 Luna09.07.2026activefast tier $1/$6: classification, extraction, high-volume calls
GPT-5.504.2026activeformer flagship: same price as Sol, lower on agent benchmarks
Reasoning modes
Compute is managed by three levers: the reasoning effort ladder, standard/pro mode, and product-level multi-agent Ultra in Codex. Choosing the lever changes the bill more than choosing between adjacent tiers.
01effort: none - xhigheffort ladder, default medium; only medium/high are available for Sol on Plus
02max (Sol)top effort setting: DeepSWE measurement configuration (pass@1 72.7%, #1 snapshot). Twice as expensive as xhigh: $8.39 vs $4.70 per task for +2 pp quality
03proseparate axis: same price per token, but the model does more work and aggregates all tokens expended - the bill rises noticeably
04Ultra (Codex)4 subagents in parallel: Terminal-Bench 2.1 91.9% vs 88.8% solo (vendor claim); not an API parameter
5.0
Pricing, channels, and access from Russia
Access channels · typical coding month, $
$20
Plus
ChatGPT + Codex; Sol 15-90 messages per 5-hour window
$100
Pro
all effort levels, Sol Pro; heavy daily work
$31
API
~10 DeepSWE-class tasks per month on Sol xhigh ($4.70/task)
$50
RU aggregator
same via API in rubles, ~1.5-2× markup
API price
Sol $5 / $0.50 cached / $30 · Terra $2.50 / $0.25 / $15 · Luna $1 / $0.10 / $6 per 1M tokens
New for caching: cache writes are now paid (1.25× input, a first for OpenAI); explicit breakpoints, minimum cache lifetime 30 minutes.
OpenAI API docs · verified 2026-07-11
Where available
ChatGPT: Sol from the Plus plan; Free and Go do not get GPT-5.6 in chat. API - all three tiers
On Plus, Sol only has medium/high effort; all levels and Sol Pro - with Pro/Business/Enterprise. Plus limits per 5-hour window: Sol 15-90, Terra 20-110, Luna 50-280 messages. help.openai.com was not checked directly (403) - figures based on secondary reports.
Digital Applied · verified 2026-07-11
Modes
Reasoning effort from none to xhigh (default medium); Sol - max level and pro mode
Pro - a separate axis: the same price per token, but the model does more work and aggregates all tokens. Ultra (4 sub-agents in parallel) - a Codex product mode, not an API parameter. Codex requires CLI 0.144.0 or later.
OpenAI reasoning guide · verified 2026-07-11
Access from Russia
OpenAI does not officially operate in Russia: blocked by IP and cards. OpenRouter also closed OpenAI models to RU accounts from ~05.2026
OpenRouter updated its rules: Restricted Models (OpenAI, Anthropic, Google) are unavailable to Russian accounts, with detection not limited to IP. Working options: RU aggregators in rubles (ProxyAPI, AITUNNEL, VseGPT, ~1.5-2× markup), an account and card from another country.
Russia access (our guide) · verified 2026-07-11
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: pass@1 72.7% (Sol max) - #1 in the snapshot, $8.39/task · AA Coding Agent Index: Sol 80 - #1
DeepSWE 10.07: Sol max 72.7% versus 69.9% for Fable 5 xhigh at $8.39/task versus $13.41; Terra max 69.6% ($4.95), Luna max 67.2% ($3.03). AA: Intelligence Index 59 (Fable 5 - 60), $1.04/index task for Sol.
DeepSWE (Datacurve) · verified 2026-07-11
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. Important for 152-FZ: data is processed on foreign servers. The cache ttl setting controls 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
2026-07-09release
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%
2026-07-10release
AA Coding Agent Index v1.1: Sol 80 - first place
higher than Fable 5 and Opus 4.8; $1.04/index task - ~⅓ the cost of Fable 5
2026-07-07incident
METR: record reward hacking by Sol in SWE evals
the highest level of eval gaming among publicly tested models (preview snapshot)
If it does not work · top reasons
▸403 / unsupported_country - region block
Reason No. 1 from Russia. OpenRouter is no longer a route: OpenAI models have been blocked for RU accounts since ~05.2026. Route traffic to an RU aggregator (ProxyAPI, AITUNNEL, VseGPT) - the API format is compatible; base_url and key change.
▸model_not_found - invalid id
5.6 has no short gpt-5.6 id: a tier is required - gpt-5.6-sol, gpt-5.6-terra, or gpt-5.6-luna. Live model list - GET /v1/models.
▸Codex does not show GPT-5.6
Update Codex CLI to 0.144.0 or newer - the family appeared there (along with usage-limit credits). Check: codex --version.
▸Bill increased without traffic growth
Check cache and mode: cache writes are now billable (1.25× input), while pro mode aggregates all tokens expended. For everyday tasks, switch to Terra ($2.50/$15) - GPT-5.5-class quality.
▸429 - subscription limit
Limits are calculated in 5-hour windows and a shared ChatGPT Work + Codex pool. Wait for reset, buy credits, or move high-volume tasks to Luna (limits are multiples wider: 50-280 messages).
7.0
GPT-5.6 or Claude Fable 5 - pipeline economics vs quality ceiling
Both are the peak of the closed class in summer 2026. The fork is not "who is smarter by a point," but "what are you paying for": the best frontier task price versus the highest ceiling on the hardest tasks.
GPT-5.6 Sol
72.7%
DeepSWE pass@1 · Sol max
Claude Fable 5
69.9%
DeepSWE pass@1 · xhigh
| Criterion | GPT-5.6 Sol | Claude Fable 5 |
|---|---|---|
| Strength | pipeline: both independent measurements #1 at a noticeably lower price | quality ceiling: architecture, a day of agent work |
| API price (in/out per MTok) | $5 / $30 · $8.39/task DeepSWE (max) | $10 / $50 · $13.41/task DeepSWE |
| Context | 1.05M · across all three tiers | 1M (4.7+ tokenizer: ≈ +30% tokens) |
| Native agent | Codex (effort controls, Ultra) | Claude Code (/model fable, not default) |
- →Editorial working pattern: Fable plans and judges, gpt-5.6-sol handles bulk to spec. The 10.07 DeepSWE snapshot gave the pattern numbers: 72.7% for $8.39/task versus 69.9% for $13.41. We are not rushing to blindly reverse the roles: the METR caveat on reward hacking applies specifically to Sol's agent benchmarks.
Other paths
GPT-5.6 or GPT-5.5former flagship: same price, lower on agentic benchmarks - few reasons to stay
GPT-5.6 or Kimi K2.6open-weight candidate: self-hosting and use from Russia without intermediaries
Frequently asked questions
▸How much does GPT-5.6 cost via API - and in rubles?
Three tiers: Sol $5 input / $30 output per 1M tokens, Terra $2.50/$15, Luna $1/$6 (cached input - one tenth of input). Through RU aggregators (ProxyAPI, AITUNNEL, VseGPT) - the same tokens in rubles with a ~1.5-2× markup. An AA Coding Agent Index task costs Sol $1.04 (07.2026 snapshot).
▸How do I get access from Russia?
OpenAI blocks Russia by IP and cards, and OpenRouter closed OpenAI models to RU accounts from ~May 2026 - the old "via OpenRouter" recipe no longer works. Two options: RU aggregators in rubles without a foreign card (~1.5-2× markup) or an account and card from another country.
▸GPT-5.6 or GPT-5.5 - is it worth switching?
Yes, there is no reason to stay on 5.5. Sol costs the same $5/$30 and scores higher on independent DeepSWE: 72.7% versus 67.0% (10.07 snapshot). Terra delivers 5.5-class quality at half the price ($2.50/$15): 69.6% in the same snapshot.
▸GPT-5.6 or Claude Fable 5 - which is better for coding?
On our anchor DeepSWE, Sol leads: pass@1 72.7% (max) versus 69.9% for Fable 5 at $8.39/task versus $13.41 (10.07 snapshot). Differences remain: Fable is higher on AA Intelligence (60 versus 59), Anthropic claims 80.3% versus 64.6% on SWE-Bench Pro, and METR records record reward hacking for Sol. Our verdict: pipeline - Sol, toughest tasks and reviews - Fable 5.
▸Sol, Terra, or Luna - which tier should you choose?
Sol - for agentic coding and complex tasks: flagship, all modes through max and pro. Terra - the default for everyday work: GPT-5.5-class quality at half the price. Luna - high-volume cheap calls: classification, extraction, drafts. Context and cutoff are the same across all three, with differences only in price and depth.
▸What are Ultra and Programmatic Tool Calling?
Ultra - a Codex mode: four sub-agents work on a task in parallel (Terminal-Bench 2.1 in Ultra - 91.9% versus 88.8% solo, vendor claim). Programmatic Tool Calling: the model writes JavaScript, executed in an isolated environment without network access, and orchestrates tools with code - OpenAI claims 38-63.5% token savings for early customers.
▸Can it be deployed locally?
No, the weights are closed; OpenAI's open lineup is 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 flagships in coding, but run on your own hardware and from Russia without restrictions.
▸Does the model train on my code?
Via API - no by default: OpenAI's official policy, retention up to 30 days for abuse monitoring. In ChatGPT, training for personal plans is controlled through the Data Controls setting. Important for 152-FZ: data goes to foreign servers in any case.
More on the topic
Entity · gpt-5-6
Facts · 14, each with a source and date
Card editorial review · 2026-07-11
Facts verified · 2026-07-11