Skip to main content

Sonnet 5.5, Opus 5.5 and GPT-6 Luna: The Cheap Tier Got Serious

· 6 min read
Gergely Sipos
Gergely Sipos
Frontend Architect

Three releases landed in one week: Opus 5.5 and GPT-6 Sol/Luna arrived in GitHub Copilot on September 22, and Sonnet 5.5 followed on September 28. I don't think the top end is the story. The story is cost: the cheap tiers now compete with last month's premium models. Every benchmark number below is vendor-reported, so read it as a claim, not a measurement.

What shipped​

ModelReleasedAPI IDInput per 1MOutput per 1M
Opus 5.5Sep 22claude-opus-5-5$4$20
Opus 5--$5$25
Sonnet 5.5Sep 28claude-sonnet-5-5$2$10
Sonnet 5--$2$10
GPT-6 SolSep 22gpt-6-sol$2$10
GPT-6 LunaSep 22gpt-6-luna$0.10$0.50

The September 22 dates are the GitHub Copilot changelog dates. Context is 1M tokens for Claude and 1.05M for OpenAI, both with 128K max output. GPT-6 Astra ($10/$50 in Copilot) is still OpenAI's best model; Sol and Luna are its cheaper siblings, and Sol is here only for context.

Claude 5.5: better and cheaper, but not the same way​

Opus 5.5 is $4/$20 against $5/$25, so about 20% cheaper per token. Anthropic claims roughly 40% cheaper than Opus 5 at default settings, because the default effort dropped from high to medium. It also claims 30%+ faster output.

Sonnet 5.5 keeps Sonnet 5's $2/$10 price. The saving comes from fewer tokens per task: Anthropic says up to 30% lower cost per task, again with 30%+ faster output.

Vendor-reported numbers:

  • Sonnet 5.5: 70.6% on Terminal-Bench 4.0.
  • Opus 5.5: 66.4% on Terminal-Bench 4.0, against 52.3% for Opus 5.
  • Opus 5.5: 54.4% on FrontierCode 1.1, against 48.0% for Opus 5.

My opinion: both models promise to beat 5 while costing less. Nice. In My Model Usage, July 2026 I admitted I skip the Sonnet rung and planned to move my middle tier onto it, so 5.5 is the obvious next test.

GPT-6 Luna: turn effort up​

Luna costs $0.10/$0.50. That is 20x cheaper than Sol and half the input price of GPT-5.6 Luna ($0.20/$1.20). Cached reads are $0.01, a 90% discount.

My recommendation, and mine alone: run Luna at max effort. It is so cheap and still pretty capable that extra reasoning tokens barely register on the bill.

The supporting evidence:

  • Effort levels are none, low, medium, high, xhigh and max, with medium as the default.
  • OpenAI's own Luna benchmarks are run at max.
  • Luna at max scores 66.6% on DeepSWE (vendor-reported). OpenAI calls that comparable to Opus 5 and Fable 5 at medium effort, at 93% and 96% lower cost. Note that those competitor baselines ran at medium.
note

This is not an official OpenAI recommendation. OpenAI's reasoning guide says to evaluate xhigh and max before adopting them. Effort can be changed mid-conversation without breaking the cache, so experimenting is cheap. More effort is also not automatically better: Anthropic reports that Sonnet 5.5 scored lower at Max than at Xhigh on FrontierCode, because subagent code-review runs caused timeouts and out-of-scope edits. Verify on your own tasks.

Read the charts carefully​

  • Every figure here comes from the vendors, not from independent testing.
  • Vendors don't benchmark their newest models against each other. Anthropic compares against GPT-5.6 Sol and Astra, and OpenAI compares against Opus 5 and Fable 5.
  • GPT-5.6 Luna and GPT-6 Luna are different models. Check the version before comparing.

Before you flip the switch​

  • Sonnet 5.5: disabled thinking, forced tool_choice and non-default temperature, top_p or top_k all return HTTP 400.
  • Opus 5.5: thinking is always on and forced tool use returns 400. Default effort is medium. Text between tool calls now lives in thinking blocks, so your UI goes quiet unless you set the display option.
  • Sol and Luna: temperature and top_p are removed when effort is not none. Function calling in Chat Completions works only with effort none, so use the Responses API instead.

See the Sonnet 5.5 migration guide and OpenAI's reasoning guide.

In GitHub Copilot​

ModelPlans
Sonnet 5.5 (GA Sep 28)Pro, Pro+, Max, Business, Enterprise
Opus 5.5Pro+, Max, Business, Enterprise
GPT-6 SolPro+, Max, Business, Enterprise
GPT-6 LunaPro, Pro+, Max, Business, Enterprise

Rollout is gradual, and Business and Enterprise admins control access through the model policy. Billing is at the provider's list price in AI credits (1 credit = $0.01). The 1M context window is only available in VS Code and Copilot CLI. Reasoning levels are configurable in VS Code, the CLI and the cloud agent.

One deadline: Opus 4.7 retires from Copilot on 2026-10-02. Opus 5 was the suggested alternative, and Opus 5.5 is now available too. For keeping the bill in check, see Chronicle cost tips and Copilot token efficiency.

My take​

Judge by cost per solved task, not per-token price, and re-run your own evals. I'd try Luna at max as the default cheap tier, and Sonnet 5.5 is the model I'll test for the middle rung. The routing argument from You Don't Deserve Fable (Yet) still applies; the rungs just got cheaper. Anthropic also promises Haiku 5.5 "in the coming weeks". For an earlier Opus launch, see Claude Opus 4.7: our first take.

Further Reading​