GPT-5.6 Sol Preview 2026: What Changed & How to Use It Now

GPT-5.6 Sol is OpenAI’s current flagship model. OpenAI launched its “dots” always-on enterprise agents this week. The demos stumbled, and the launch landed while OpenAI is reviewing agent incidents that affected more than 100 organizations. That makes this a good time to ask what Sol actually changes, who can access it, and whether moving your workflows to it is worth the effort. Our short answer: Sol is a real upgrade for long, agentic work, but test it on your own tasks before switching everything.

What’s actually new in GPT-5.6 Sol

OpenAI previewed GPT-5.6 Sol on June 26, 2026, alongside two siblings: GPT-5.6 Terra, a balanced model for everyday work, and GPT-5.6 Luna, a fast, low-cost model for high-volume jobs. At first, only a small group of vetted partners could use Sol. The US government requested that restriction, and OpenAI publicly disagreed with holding the model back. The whole family became generally available across ChatGPT, Codex, and the API on July 9, 2026.

The headline changes are about how much work the model can do in one go. Sol adds a new max reasoning effort at the top of the scale. It also adds an ultra mode that splits complex work across sub-agents running in parallel. According to the official model page, the API model has a 1,050,000-token context window, up to 128,000 output tokens, and a knowledge cutoff of February 16, 2026. OpenAI says Sol improves agentic coding, biology, and cybersecurity work. Third-party benchmark write-ups report 88.8% on Terminal-Bench 2.1 for Sol, 91.9% in ultra mode, and 88.0% for GPT-5.5.

The cybersecurity results explain the cautious rollout. OpenAI reports that Sol found real bugs in Chromium and Firefox but did not produce a full working exploit on its own, which keeps it below the company’s “Cyber Critical” threshold. On ExploitBench, it reportedly matched a leading competitor while using about a third of the output tokens. OpenAI calls Sol’s safety stack its “most robust to date,” which is why some security-sensitive requests get stricter handling than they did on GPT-5.5.

Why it matters

• Long tasks get more reliable. A million-token context, 128K-token outputs, and max effort let Sol keep a whole codebase or a large document set in view and still produce a complete result. You need fewer manual handoffs.
• Ultra mode changes the cost math. Parallel sub-agents finish hard tasks faster, but every sub-agent uses tokens. Turn ultra on for specific jobs; don’t make it your default.
• The flagship costs less. Sol launched at $5 input and $30 output per million tokens. It is now listed at $4 and $20, with cached input at $0.40. Third-party trackers say this promotional pricing runs at least through November 21, 2026.
• Terra is the quiet winner for most teams. OpenAI says Terra matches GPT-5.5 at about half the cost. If your workload already ran fine on GPT-5.5, make Terra your first migration target, not Sol.
• Governance is now part of choosing a model. The government held back Sol’s launch, and OpenAI’s agents are under review this week. Enterprise buyers should expect stricter policies, more logging requirements, and closer audits of anything agentic.
• The competitive gap is narrow. Reported Terminal-Bench numbers put Sol less than a point ahead of GPT-5.5 and Anthropic’s top model. Ultra mode shows the bigger jump. Choose a model based on your own task results, not leaderboard headlines.

How to access GPT-5.6 Sol and use it today

1. Check your ChatGPT plan

   Sol is available on Plus and above, including Pro, Business, and Enterprise, at medium and higher effort settings. Pro and Enterprise also get GPT-5.6 Sol Pro for the hardest tasks. Open the model picker, choose GPT-5.6 Sol, and raise the effort setting for multi-step work. Options vary by plan; if something is missing, check OpenAI’s GPT-5.6 help article.

2. Call it from the API

   The model ID is gpt-5.6-sol. It works on the Responses, Chat Completions, and Batch endpoints. Start with the default effort and raise it one level at a time only where quality demands it.

   from openai import OpenAI
   
   client = OpenAI()
   
   response = client.responses.create(
       model="gpt-5.6-sol",
       reasoning={"effort": "high"},
       input="Review this migration plan for data-loss risks and list fixes in priority order:\n\n<paste plan here>"
   )
   
   print(response.output_text)

   The same call with curl:

   curl https://api.openai.com/v1/responses \
     -H "Authorization: Bearer $OPENAI_API_KEY" \
     -H "Content-Type: application/json" \
     -d '{
       "model": "gpt-5.6-sol",
       "reasoning": {"effort": "medium"},
       "input": "Summarize the key risks in this contract clause: ..."
     }'

3. Use the effort ladder on purpose

   The supported values are none, low, medium (the default), high, xhigh, and max. Here’s how we map them:

   # Rough effort-to-task mapping (start low, escalate on failure)
   none / low   -> classification, extraction, short rewrites
   medium       -> drafting, standard code changes, Q&A over docs
   high / xhigh -> multi-file refactors, analysis with tradeoffs
   max          -> long-horizon agent runs, research, hard debugging

4. Set it as your default in Codex

   In Codex, Plus and higher plans can choose Sol, Terra, or Luna, set an effort level per model, and turn on ultra mode. In the CLI, set it in your config:

   # ~/.codex/config.toml
   model = "gpt-5.6-sol"
   model_reasoning_effort = "high"

   Some developers report that Codex rejects gpt-5.6-sol when signed in with a ChatGPT account instead of an API key. If you hit that error, sign in with an API key or pick the model from the Codex model picker.

5. Push bulk work through Batch

   For evals, backfills, and content jobs that don’t need an immediate answer, the Batch API roughly halves the cost. Write one request per line in a JSONL file:

   {"custom_id": "doc-001", "method": "POST", "url": "/v1/responses", "body": {"model": "gpt-5.6-sol", "reasoning": {"effort": "low"}, "input": "Tag this support ticket by product area: ..."}}

6. Structure prompts around goals

   Sol performs best when you give it a goal, constraints, and a clear definition of done instead of step-by-step instructions:

   Goal: Reduce p95 latency of the /search endpoint below 300ms.
   Constraints: No schema changes. Keep the public API identical.
   Context: [attach profiler output + relevant files]
   Definition of done: A patch, a short explanation of each change,
   and a test plan I can run locally.

7. Cache your long prompts

   Cached input costs a tenth of the normal rate, so put stable content first (system prompt, policies, reference docs) and variable content last. Watch request size too: third-party pricing guides report a surcharge on prompts over about 272K tokens, so a full million-token context gets expensive fast.

How GPT-5.6 Sol compares to GPT-5.5 and its siblings

Model	Best for	Standout features	Reported price (per 1M tokens)
GPT-5.6 Sol	Long-horizon agents, hard coding, research	max effort, ultra sub-agent mode, 1.05M context, 128K output	$4 input / $20 output (promo; launched at $5 / $30)
GPT-5.6 Terra	Everyday production workloads	About GPT-5.5-level quality at about half the cost	About $2.50 / $15 at preview
GPT-5.6 Luna	High-volume, latency-sensitive jobs	Lowest cost in the family	About $1 / $6 at preview
GPT-5.5	Existing pipelines you haven’t re-tested yet	Mature, well understood; reported 88.0% on Terminal-Bench 2.1 vs Sol’s 88.8%	Being displaced by Terra on price/performance

Our view: for most teams, moving from GPT-5.5 to Terra is the easy call, since you keep similar quality at lower cost. Move to Sol only for workloads where GPT-5.5 actually fell short, such as long agent runs, large refactors, or deep analysis. That’s where max effort and ultra mode earn their cost.

What’s next

Watch speed first. OpenAI has previewed an Ultrafast service tier, powered by Cerebras hardware, that runs GPT-5.6 Sol up to 14 times faster than standard processing, at up to roughly 750 output tokens per second. If it reaches general availability at a reasonable price, it changes what’s practical for interactive agents and coding loops, many of which are limited today by waiting time, not model quality.

Watch governance second. Sol was the first OpenAI flagship whose release was held back at the government’s request, and OpenAI is working with officials on a framework for future releases under a cyber executive order. Add the current review of agent incidents and the shaky dots demos, and expect more paperwork around agentic deployments, not less: admin controls, audit logs, and narrower default permissions. Plan for them now so they don’t block you later.

Finally, watch pricing. The current Sol rates are promotional, and third-party trackers say they last at least through late November. Before you budget a large migration, model your costs at both the current rates and the original $5/$30 launch pricing.

Frequently Asked Questions

Is GPT-5.6 Sol still in preview?

No. The limited partner preview started June 26, 2026, and Sol, Terra, and Luna became generally available across ChatGPT, Codex, and the API on July 9, 2026. Some features, such as the Ultrafast tier, remain in preview.

How do I access GPT-5.6 Sol?

In ChatGPT, choose it from the model picker on a Plus, Pro, Business, or Enterprise plan. In the API, use the model ID gpt-5.6-sol with the Responses, Chat Completions, or Batch endpoints. In Codex, Plus and higher plans can select it and set the reasoning effort.

What is ultra mode?

Ultra mode splits complex work across several sub-agents running in parallel. In Codex it’s available from Plus upward; in ChatGPT Work it’s limited to Pro and Enterprise. It finishes hard tasks faster but uses noticeably more tokens, so save it for tasks that need it.

Should we switch from GPT-5.5 to GPT-5.6 Sol now?

Only for workloads where GPT-5.5 fell short. For general production work, test Terra first; it reportedly matches GPT-5.5 at about half the cost. Run your own evals on real tasks before switching anything.

Why was the GPT-5.6 Sol launch restricted?

Sol is strong at long-horizon cybersecurity work such as vulnerability research, so the US government asked OpenAI to limit the first preview to vetted partners. OpenAI said the model did not cross its “Cyber Critical” threshold and opened access broadly about two weeks later.

How much does GPT-5.6 Sol cost?

The current API price is $4 per million input tokens and $20 per million output tokens, with cached input at $0.40, down from $5 and $30 at launch. Batch requests cost less, and very long prompts carry a surcharge.