Loading the page.
Card · LLM for code · updated 2026-07-11
GPT-5.5
OpenAI · flagship agent · reasoning up to xhigh · 1.05M context · closed weights
67.0%
DeepSWE pass@1 · independent measurement
#4out of 16 models
Editorial review 2026-07-11
Editorial review 2026-07-11
OpenAI's flagship model for agentic coding: handles long multistep tasks, takes them through diff and validation, and saves tokens vs. its predecessors. On DeepSWE in top xhigh mode - pass@1 67.0%: behind GPT-5.6 Sol (72.7%) and Fable 5 (69.9%). Since 09.07, it has had a successor at the same price - GPT-5.6.
$5/$30
api per 1M tokens · cached input $0.50
1.05M
api · 400k in codex
12.2025
model knowledge cutoff
Yes
Direct payment from Russia
from Russia: officially unavailable · openrouter or an aggregator in rubles
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
- +complex PRs, architecture, debugging, and multi-file changes
- +tasks where the cost of an error exceeds token costs
- +top quality baseline: compare cheaper models against it
Do not use for
- −mass simple edits - overpaying; gpt-5.4-mini or a cheaper tier is enough
- −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 (see Connection)
Until 09.07, a ChatGPT subscription (Plus $20, Pro from $100) included GPT-5.5 in ChatGPT and Codex; since 09.07, OpenAI's primary model is GPT-5.6. GPT-5.5 via API - $5/$30 per 1M tokens. From Russia - via OpenRouter or RU aggregators.
2.0
Access and payment
Codex
codex /model gpt-5.5
In Codex, GPT-5.5 was the primary model until 09.07; it was replaced by GPT-5.6.
OpenAI API
curl https://api.openai.com/v1/responses \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-d '{"model": "gpt-5.5", "input": "..."}'Canonical id - gpt-5.5; reasoning effort is set by a parameter.
OpenRouter
curl https://openrouter.ai/api/v1/chat/completions \
-H "Authorization: Bearer $OPENROUTER_KEY" \
-d '{"model": "openai/gpt-5.5", "messages": [...]}'Same $5/$30 price; pay with crypto or a foreign card.
RU aggregator
curl https://api.proxyapi.ru/openai/v1/chat/completions \
-H "Authorization: Bearer $PROXYAPI_KEY" \
-d '{"model": "gpt-5.5", "messages": [...]}'OpenAI-compatible format: only base_url and the key change. Rubles, no VPN, ~1.5–2× markup.
One brain - different names · model ID by surface
| OpenAI API | gpt-5.5 | canonical |
| Codex | gpt-5.5 | primary until 09.07; select via /model |
| OpenRouter | openai/gpt-5.5 | |
| Cursor | gpt-5.5 | in model picker |
| RU aggregators | gpt-5.5 | ProxyAPI / AITUNNEL / VseGPT |
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. Working routes: OpenRouter (USDT crypto, foreign card), RU aggregators in rubles without VPN - ProxyAPI, AITUNNEL, VseGPT (~1.5–2× markup). DeepSeek is the only strong option that works directly from Russia and is free.
3.0
Ranking position and benchmark variation
67.0%DeepSWE pass@1 · 07.2026independent measurement by deepswe.datacurve.ai; GPT-5.6 Sol max - 72.7%, Fable 5 - 69.9%
82.7%Terminal-Bench 2.0 · 04.2026vendor claim from the release
58.6%SWE-bench Pro · 04.2026here, Opus 4.8 leads with 69.2% - benchmarks diverge
Benchmarks differ - that is normal
Benchmarks have split, and that should be read honestly: Claude Opus 4.8 leads on SWE-bench Pro, GPT-5.5 on Terminal-Bench, and GPT-5.6 Sol has led DeepSWE since 10.07 (72.7%). The old SWE-bench Verified is contaminated and no longer distinguishes flagships - so we show the variation across independent measurements, not one convenient number.
model ≠ agent
Model score ≠ tool score: GPT-5.5 was tested until 09.07 both paired with Codex (agent benchmarks) and solo (model benchmarks). Compare identical configurations - DeepSWE fixes xhigh mode.
4.0
Family, versions, and modes
A model is a versioned series, not a point: OpenAI flagships update every 6–10 weeks. The card tracks the GPT-5.x family; stale versions move to deprecated.
GPT-5.6 (Sol/Terra/Luna)09.07.2026flagshipnew three-tier generation: Sol - the same $5/$30, stronger on agent benchmarks
GPT-5.504.2026activeagentic coding, computer use; replaced by GPT-5.6 as flagship since 09.07
GPT-5.5-pro04.2026activetop $30/$180 per MTok tier - 6× more expensive, almost always excessive for code
GPT-5.4 / 5.4-mini02.2026activecheaper ($2.50/$15 and $0.75/$4.50): workhorse tier for simple tasks
GPT-5.2, GPT-5.3-codex2025deprecatedthere is no separate “-codex” line anymore - consolidated into gpt-5.5
Reasoning modes
Within one model - a reasoning effort ladder: the higher it is, the longer it thinks and the more it costs. Mode selection changes both quality and the bill more than choosing between adjacent versions.
01low / mediumfast answers and simple edits; save tokens, do not handle complex debugging
02highworkhorse mode for everyday agentic coding
03xhighmaximum quality on complex tasks - the DeepSWE measurement configuration (pass@1 67.0%). More expensive and slower: enable where the cost of an error is high
04gpt-5.5-proseparate pro tier $30/$180: research-class tasks, almost always excessive for coding
5.0
Pricing, channels, and access from Russia
Access channels · typical coding month, $
$20
Plus
ChatGPT + Codex; limits in 5-hour windows
$100
Pro
limits ×5–20; heavy daily work
$33
API
~5 DeepSWE-class tasks per day ($6.61/task)
$55
RU aggregator
the same via API in rubles, ~1.5–2× markup
API price
$5 input / $0.50 cached / $30 output per 1M tokens
Above 272K input tokens in the prompt, the price rises: ×2 input and ×1.5 output for the entire session; data-residency-endpoints +10%.
OpenAI API docs · verified 2026-07-03
Where available
ChatGPT and Codex: Go, Plus, Pro, Business, Enterprise, Edu plans; API - metering
In Codex, it was the primary model until 09.07 (replaced by GPT-5.6); limits - messages per 5-hour window (Plus: 15-80 on gpt-5.5), with credit purchases beyond that.
Codex models · verified 2026-06-10
Modes
Reasoning effort from low to xhigh; DeepSWE measures xhigh
A separate top-tier gpt-5.5-pro costs $30/$180 per MTok - 6 times more expensive; before the 09.07 transition, gpt-5.5 xhigh was the primary choice for coding.
OpenRouter API · verified 2026-07-03
Access from Russia
OpenAI does not officially operate in Russia: blocked by IP and cards
Working options: OpenRouter (crypto/foreign card), Russian aggregators in rubles with a ~1.5-2× markup (ProxyAPI, AITUNNEL, VseGPT). A ChatGPT subscription cannot be paid for directly from Russia.
Russia access (our guide) · verified 2026-07-02
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
Terminal-Bench 2.0 - 82.7%; SWE-Bench Pro - 58.6%
Figures from the OpenAI release are vendor claims, not independent measurements; the independent snapshot is below.
OpenAI release · verified 2026-06-05
Independent snapshot
DeepSWE: pass@1 67.0%, pass@4 88.5%, $7.23/task
Independent measurement by deepswe.datacurve.ai (07.2026 snapshot); leader of previous snapshots, since 10.07 GPT-5.6 Sol is first (72.7%), Fable 5 second (69.9%). Artificial Analysis Coding Index - 76.4.
DeepSWE (Datacurve) · verified 2026-07-08
Data
API traffic is not used for training by default; in ChatGPT - depends on settings
Official OpenAI policy for API/Business: customer data is not used for training without opt-in. Important for 152-FZ: data is processed on foreign servers.
OpenAI enterprise privacy · verified 2026-07-02
Weights
Closed weights - cannot be deployed locally
Open-weight alternatives for code - Kimi K2.6, GLM-5.x; they lag behind flagships in coding quality.
OpenAI API docs · verified 2026-07-03
Track record · recent events
2026-04-23release
GPT-5.5 release: agentic coding, computer use, knowledge work
Terminal-Bench 2.0 82.7%, SWE-Bench Pro 58.6% according to the vendor claim
2026-06-10policy
Lineup consolidation: gpt-5.2 and gpt-5.3-codex - deprecated
there is no separate flagship "-codex" branch anymore; gpt-5.5 - Codex default
2026-06-04release
DeepSWE snapshot: pass@1 70.0% - leader of the June measurement
independent benchmark deepswe.datacurve.ai; in the 10.07 snapshot GPT-5.6 Sol is first (72.7%)
If it does not work · top reasons
▸403 / unsupported_country - region block
Reason #1 from Russia. Route traffic through OpenRouter (openai/gpt-5.5) or an RU aggregator - the API format is compatible, only base_url and the key change.
▸429 / 529 - overload or rate limit
Check status.openai.com; reduce concurrency for batches. On a subscription - a 5-hour window limit: wait for reset or buy additional credits.
▸Context overflowed, auto-compact consumes details
Above 272K input, the price is ×2 - give the model an exact slice, not the whole repo. In Codex, context is 400K: split the task into stages.
▸model_not_found - invalid id
Check the id against the surface table: gpt-5.5 in API and Codex, openai/gpt-5.5 on OpenRouter. List of live models - GET /v1/models.
▸“The model got dumber” after a silent update
Check whether the snapshot and reasoning mode changed (low effort feels like degradation). Pin the version in the API request and compare on the same task.
7.0
GPT-5.5 or Claude Opus 4.8 - the flagship fork of spring 2026
Both are closed-class flagship agents. The fork is not “who is smarter,” but “which pipeline”: cheap mass grind in Codex versus depth on complex context in Claude Code.
GPT-5.5
67.0%
DeepSWE pass@1 · xhigh
Claude Opus 4.8
60.9%
DeepSWE pass@1 · max
| Criterion | GPT-5.5 | Claude Opus 4.8 |
|---|---|---|
| Strength | agentic pipeline: long tasks, token savings | large codebase architecture, precise edits |
| API price (in/out per MTok) | $5 / $30 | $5 / $25, but token usage is higher |
| Context | 1.05M API / 400K Codex | 1M |
| Native agent | Codex (default model) | Claude Code |
- →The 2026 norm - keep both subscriptions and route tasks by type: quality is near parity, but the genres differ.
Other paths
GPT-5.5 or GPT-5.6successor: the same $5/$30 for Sol, stronger on agent benchmarks; Terra is half price
GPT-5.5 or Kimi K2.6open-weight candidate: self-hosting and use from Russia without intermediaries
Frequently asked questions
▸How much does GPT-5.5 cost via API - and in rubles?
Officially $5 input / $30 output per 1M tokens (cached input $0.50). Through Russian aggregators (ProxyAPI, AITUNNEL, VseGPT) - the same tokens in rubles with a ~1.5-2× markup. A typical Codex task in our DeepSWE snapshot - $6.61.
▸How do I get access from Russia?
OpenAI blocks Russia by IP and cards. Three options: OpenRouter (payment with crypto or a foreign card, slug openai/gpt-5.5), Russian aggregators in rubles without VPN (~1.5-2× markup), an account and card from another country. A ChatGPT subscription requires the second or third option.
▸GPT-5.5 or Claude Opus 4.8 - which is better for code?
The benchmarks honestly disagree: Opus leads on SWE-bench Pro (69.2% vs. 58.6%), while GPT-5.5 leads on our anchor DeepSWE (70.0% vs. 58%). Successor GPT-5.6 Sol max - 72.7% DeepSWE vs. 70.0% for GPT-5.5 in the June snapshot. GPT-5.5 is 2 times cheaper than Opus on output and more token-efficient. Our verdict: GPT-5.6 for the current Codex pipeline, Opus for complex architecture in Claude Code.
▸What does xhigh mean and when is it needed?
It is the top reasoning effort: the model thinks longer and costs more, but is noticeably stronger on complex tasks. DeepSWE measures xhigh specifically. For minor edits, lower modes or a cheaper model are enough - there is no reason to pay the xhigh price for renaming a variable.
▸How much code fits in the context?
1.05M tokens via API - roughly 3-4 MB of source code, an entire average repository. But above 272K input, the price doubles, and Codex has a 400K context: in practice, it is more cost-effective to give the model a precise slice rather than the entire repo.
▸Subscription or API - which is more cost-effective?
If you work in Codex daily - a ChatGPT subscription (Plus $20, Pro from $100) is almost always cheaper than metering: limits are counted in messages, not tokens. The API is better for automations, CI, and when you need to control spending by task.
▸Can it be deployed locally?
No, the weights are closed. If you need self-hosting - look at the open-weight class: Kimi K2.6, GLM-5.x, DeepSeek V4. They lag behind GPT-5.5 in code, but run on your own hardware and from Russia without restrictions.
▸Does the model train on my code?
Via API - no by default (official OpenAI policy for API and Business plans). In ChatGPT personal plans, training is managed through the Data Controls setting. Important for 152-FZ: data is sent to foreign servers in any case.
More on the topic
Entity · gpt-5-5-extra-high
Facts · 12, each with a source and date
Card editorial review · 2026-07-11
Facts verified · 2026-07-03