GPT-5.6 Sol: The Flagship Model of Summer 2026

Status as of November 5, 2026. GPT-5.6 Sol is no longer OpenAI’s most capable model: GPT-6 Astra surpassed it on September 3, and GPT-6.1 Sol has offered performance close to Astra since September 29 at $2 / $10 per million tokens (Lesson 9). GPT-5.6 Sol remains available via the API and serves as the official replacement for gpt-5 and o3, which were retired on December 11. Its price dropped from $5 / $30 to $4 / $20 on August 21, 2026, a promotional rate the pricing page lists through November 30, 2026.

On July 9, 2026, after a two-week limited preview, OpenAI made the GPT-5.6 family available in ChatGPT, Codex, and the API. Sol is the flagship model, positioned as the reference reasoning engine for complex work: coding, research, science, cybersecurity, computer control, and design. This lesson describes what it brought to the table and how to call it. This information remains useful: many projects still rely on it, and it officially replaces several end-of-life models.

Technical Specifications (Recorded November 5, 2026)

Two lines in this spec sheet deserve explanation. The alias gpt-5.6 points to Sol: if your code simply requests gpt-5.6, the flagship model responds at the flagship rate. To control costs with Terra or Luna, you must name them explicitly; the next lesson covers this. Additionally, the documentation notes that Sol roughly corresponds to the suffix-less model in previous GPT-5 families. This is a solid migration benchmark: where you used gpt-5, Sol is the natural successor, and indeed the one recommended on the deprecations page.

What OpenAI Highlighted

According to the July 9 announcement, GPT-5.6 was trained to extract more useful work from every token. This was the guiding theme of the entire presentation: not “smarter in the abstract,” but more efficient per dollar spent. OpenAI claimed state-of-the-art results on a long-horizon professional work evaluation covering 55 domains, Agents’ Last Exam, and on several coding benchmarks: the Artificial Analysis Coding Agent Index, Terminal-Bench 2.1, and DeepSWE. All this, according to the publisher, with fewer output tokens and a lower estimated cost than competing models.

As always with results published by a vendor on their own models, exercise caution. Focus on the trend—the priority given to per-token efficiency—rather than the rankings themselves, and verify on your own use cases before drawing conclusions. A model that dominates Terminal-Bench may disappoint on your internal codebase, and vice versa: only an evaluation on your data settles the matter.

In ChatGPT, At Launch

In July, Plus, Pro, Business, and Enterprise plans accessed GPT-5.6 Sol in Chat mode starting at medium effort, while Pro and Enterprise could select GPT-5.6 Sol Pro for the most complex tasks. In ChatGPT Work and Codex, Free and Go plans used GPT-5.6 Terra, while paid plans chose between Sol, Terra, and Luna with an effort level. This landscape has since changed: Chat mode has relied on GPT-6 since October 7, and Work and Codex now offer GPT-6 models (Lessons 9 and 11).

A Basic Call

The call below submits a contract for analysis with intermediate reasoning effort:

from openai import OpenAI

client = OpenAI()

response = client.responses.create(
    model="gpt-5.6-sol",
    input="Analyze this contract and list unusual clauses.",
    reasoning={"effort": "medium"},
)

print(response.output_text)

The reasoning parameter is part of the call itself. For routine contract analysis, medium suffices; if the contract is dense and missing a clause is costly, bump it to high or beyond without changing anything else. For the most difficult tasks that tolerate higher latency, add "mode": "pro" within that same object: the model then performs more work, billed at its standard token rate, but consumes more tokens.

Ultra: A ChatGPT Setting, Not an API Parameter

The July announcement also introduced ultra, the maximum capacity setting. This is not just another notch on the effort slider: ultra coordinates four agents in parallel by default to finish complex tasks faster, and OpenAI demonstrated configurations with sixteen agents on certain benchmarks. The trade-off is explicit: significantly higher token consumption. At launch, ultra was available in ChatGPT Work for Pro and Enterprise plans, and in Codex starting with Plus.

In the API, there is no ultra value. OpenAI states that developers can build comparable experiences using the multi-agent function of the Responses API, currently in beta: the model launches sub-agents in parallel and synthesizes their work in a single request. Lesson 4 returns to these settings.

Your Turn

Verify the alias and reasoning cost. If you have API access, send the same request twice, once with model="gpt-5.6" and once with model="gpt-5.6-sol", then compare the model field in both responses. Next, rerun the request with low effort, then high, and compare the reasoning token count reported in usage. Without API access, review the technical spec sheet and note the two lines that would make you prefer GPT-6.1 Sol for a new project.

Official Sources: GPT-5.6 Announcement and Preview Announcement (OpenAI) · GPT-5.6 Sol Model Card · Pricing (OpenAI API Documentation) · Model Release Notes (OpenAI Help Center)

Key Takeaways

• gpt-5.6-sol, aliased as gpt-5.6, was the July flagship; GPT-6 Astra and GPT-6.1 Sol surpassed it in the fall
• Effort ranges from none to max (medium default), pro mode, 1.05 million token context
• $4 / $20 per million tokens through November 30, 2026, down from $5 / $30 at launch
• Ultra, defaulting to four parallel agents, is a ChatGPT Work and Codex setting; in the API, use the multi-agent beta
• Verify publisher claims on your own use cases before deciding