본문으로 건너뛰기

GPT-5.6 & ChatGPT Work for Claude Users

중급

You already think in Claude. Then on 9 July 2026 the whole timeline flipped: OpenAI shipped three GPT-5.6 variants — Sol, Terra, Luna — and ChatGPT Work, an agent that takes a goal and works on it in the background for hours. A teammate asks "should we swap to Sol?"; a client wants ChatGPT Work to replace half a workflow you already run in Claude. This page maps the launch onto your Claude mental model and — more usefully — pulls out the parts the news coverage buries: the real prices, the exact benchmark deltas, the 1.05M context nobody expected, and where OpenAI still loses to Anthropic.

What you'll learn
  • Learn what actually shipped: Sol, Terra, Luna (plus Sol Ultra), all with a 1.05M-token context window
  • Read the benchmark table the way OpenAI didn't publish it — where GPT-5.6 leads and where Claude Fable 5 still wins
  • Understand ChatGPT Work: not a chat mode, but a background agent that returns a finished spreadsheet, deck, doc, or web app hours later
  • Know when to reach for GPT-5.6 Sol over Claude for coding, agents, browser tasks — and when the answer is still Claude

The one-sentence version​

GPT-5.6 is a three-tier refresh of the GPT-5 family — Sol (flagship), Terra (mid), Luna (budget) — sharing a 1.05M-token context window, 128K max output, and an OpenAI-published claim of 54% better token efficiency on agentic coding; ChatGPT Work is the new autonomous agent built on top of it, delivered as a merged desktop app that swallowed Codex.

The interesting part isn't the model refresh — that was expected. It's the shape of the release: OpenAI is now shipping model + agent + IDE as one product, and pricing the workhorse tier ($1 in / $6 out on Luna) below where "cheap coding models" have historically sat.

Three things about the GPT-5.6 launch that surprise people​

Everyone's writing the benchmark recap. Here are the parts most Claude users miss on the first read.

1. All three tiers share the same 1.05M context​

This is not how OpenAI usually ships. In earlier refreshes, the cheap tier came with a shrunken context window as a way to nudge you upward. In GPT-5.6, Sol, Terra, and Luna all quote a 1.05M-token context and 128K max output — the same shape. That means Luna at $1 input / $6 output per million tokens is a viable long-context option, not just a short-request one, and it changes the maths for RAG pipelines, whole-repo review, and long tool loops. If you were budgeting a Claude Haiku-style tier for high-volume retrieval work, Luna is now the obvious cross-check.

2. ChatGPT Work is not "a chat mode" — it's a background agent that hands you a finished file​

The framing OpenAI uses is telling: you don't chat with ChatGPT Work, you brief it. It takes an outcome ("build a competitive analysis of these five vendors"), gathers context from your connected apps and files, breaks the job into steps, works independently for minutes or hours, and delivers a finished artifact — a spreadsheet, slide deck, document, report, or working web app. Codex was folded into the same desktop app, so ChatGPT Work also owns the code-writing surface. This is the same "long-running autonomous work" bet Anthropic made with Claude's managed agents, but shipped as an end-user product first and an API second. It rolled out on Pro / Enterprise / Edu on launch day, with Plus and Business following within days.

3. Sol wins the OpenAI-published benchmarks. Claude still leads SWE-bench Pro.​

Read past the marketing chart and the picture is more mixed than "Sol is the new king". OpenAI's own numbers say Sol reaches 80 on the Artificial Analysis Coding Agent Index (about +2.8 vs Claude Fable 5) while using less than half the output tokens at roughly a third the cost. That's a real win. But Fable 5 still leads SWE-Bench Pro by a large reported margin — Sol comes in at ~64.6% there — and the launch coverage tends to bury that. Translation: pick by workload, not by press release. For "agent index" style tool-and-browser work, Sol is now the model to beat. For real-repo bug fixes with meaningful context, Anthropic still has the edge.

The GPT-5.6 tier table (as of 9 July 2026)​

ModelInput $/1MOutput $/1MContextMax outputPositioning
Sol$5$301.05M128KFlagship: complex reasoning, coding, science, agents
Terra$2.50$151.05M128K"GPT-5.5 performance at ~half the cost" — routine work
Luna$1$61.05M128KHigh-volume, simpler tasks; long-context on a budget
Sol Ultra(not disclosed)(not disclosed)1.05M128KHigher-effort Sol tier used in some published benchmarks; also the surface that exposes the multi-agent Responses beta

Benchmarks OpenAI published (Sol / Terra / Luna):

  • Artificial Analysis Coding Agent Index v1.1: 80 / 77.4 / 74.6
  • SWE-Bench Pro: 64.6% / 63.4% / 62.7% (Claude Fable 5 still leads this one)
  • Terminal-Bench 2.1: 88.8% (Sol Ultra 91.9%) / 87.4% / 84.7%
  • DeepSWE v1.1: 72.7% / 69.6% / 67.2%
  • BrowseComp: 90.4% (Sol Ultra 92.2%) / 87.5% / 83.3%
  • ExploitBench: 73.5% / 52.9% / 33.2%

That last row — ExploitBench — is worth staring at. The Sol → Luna drop from 73.5% to 33.2% is much sharper than any other benchmark on the list, which is why OpenAI leaned on the cybersecurity story for Sol specifically and why federal reviewers held the model for extra testing before letting it out.

ChatGPT Work, decoded​

Enter 또는 스페이스 키를 눌러 카드를 뒤집습니다. 좌우 화살표 키로 카드를 이동할 수 있습니다.용어가 표시되었습니다.
1 / 6

What transfers from Claude (most of it)​

Your prompting instincts still work. Clear instructions, explicit output shapes, few-shot examples, and giving the model tools with good descriptions — all of it moves across, because these are properties of instruction-tuned frontier models, not Claude-specific tricks. For the deep version of that argument see porting prompts across models and the practical guide ChatGPT for Claude users.

Two things to reset when you jump:

  • API shape. Anthropic uses a tools block on the messages API; OpenAI uses tools on Chat Completions (or the Responses API), with the OpenAI function-calling schema. Same idea, different wire format.
  • Reasoning as a mode, not just a slider. Claude exposes a thinking/effort parameter; on GPT-5.6, the sharpest gains show up when you pick the right tier (Sol / Sol Ultra) rather than only tuning effort on the cheap one. Cost matters more than it did on GPT-5.5.

Try it in a few lines​

The OpenAI SDK unchanged — just point model at a GPT-5.6 variant:

Call GPT-5.6 Sol from Python (OpenAI SDK)

from openai import OpenAI

client = OpenAI()  # uses OPENAI_API_KEY

resp = client.chat.completions.create(
  model="gpt-5.6-sol",         # try gpt-5.6-terra / gpt-5.6-luna to trade cost for quality
  messages=[
      {"role": "system", "content": "You are a careful research agent. Use tools; verify before you conclude."},
      {"role": "user", "content": "Compare our 3 largest competitors on pricing, packaging, and 2026 launches."},
  ],
  # tools=[...]  # OpenAI-style function/tool schema
)

print(resp.choices[0].message.content)
Guided walkthrough1 of 4
  1. Sol for complex coding, agent loops, or anything where you'd have reached for a Claude Opus tier. Terra as the default 'GPT-5.5 at half the cost' workhorse. Luna for retrieval, extraction, classification, and other high-volume jobs — the shared 1.05M context makes it a real option, not a toy.

When to reach for GPT-5.6 over Claude​

Neither is strictly better. Reach for GPT-5.6 when:

  • You need a $1/$6 tier with a 1.05M context. Luna is currently the cheapest way to get long-context frontier-adjacent quality; that's a new capability, not a rebrand.
  • The task is browser-heavy or agent-index-shaped. BrowseComp and the Coding Agent Index are Sol's strongest cards. If your agent spends most of its time in a browser or a shell, Sol is the model to beat.
  • You want ChatGPT Work's "brief and forget" workflow for producing finished spreadsheets/decks/reports/apps as end-user artifacts, without gluing an agent framework yourself.

Reach for Claude when you want the managed platform around the model — prompt caching, memory and context editing, managed agents, computer use, the desktop app — and specifically when the workload is real-repo software engineering, where Fable 5 still leads SWE-Bench Pro by a large margin. See Claude, GPT, Gemini for coding for the ongoing head-to-head, and what AI costs across providers for the cost view.

Check yourself

0/3
  1. Which GPT-5.6 tier has the smallest context window?
  2. What is ChatGPT Work, in one line?
  3. On which OpenAI-published benchmark does Anthropic's Claude Fable 5 still lead GPT-5.6 Sol?

Sources & further reading​