GPT-6 Luna vs Claude Sonnet 5.5: Two Gaps, Not One

Put GPT-6 Luna and Claude Sonnet 5.5 on the same price sheet and the ratio is embarrassing: $0.10 against $2.00 per million input tokens, $0.50 against $10.00 per million output. Put them on the same evaluation and the gap looks completely different depending on which evaluation you pick. On Artificial Analysis's Intelligence Index, Sonnet 5.5 at maximum effort scores 55 and Luna at maximum effort scores 28 - a 27-point spread on a board where the top of the field sits in the fifties. On FrontierCode-AA 2.0, the same two models score 25.7 and 25.2. That is not a gap. That is a tie.

Both readings are correct, and the reason they coexist is the thing worth understanding before choosing. Luna is not a uniformly weaker model selling at a twentieth of the price. It is a model with a specific shape: competitive on code-diff generation, far behind on terminal work, scientific reasoning and factual restraint, and roughly an order of magnitude slower to produce its first token. Sonnet 5.5 is the opposite shape - consistent, quick to start, and expensive enough that a fleet of them needs a business case.

The two spec sheets, dimension by dimension

• Input price per 1M - GPT-6 Luna $0.10 vs Claude Sonnet 5.5 $2.00, a 20x ratio

• Output price per 1M - GPT-6 Luna $0.50 vs Claude Sonnet 5.5 $10.00, also 20x

• Cached input per 1M - GPT-6 Luna $0.01 vs Claude Sonnet 5.5 $0.10; A​nthropic cut Sonnet's cache-read rate to half of Sonnet 5's at launch, and the ratio to Luna's cached rate is still 10x

• Cache writes - GPT-6 Luna $0.125 per 1M, a 1.25x premium on input; Sonnet 5.5 is billed at $2.50 per 1M on the same 1.25x convention

• Context window - 1,050,000 tokens for GPT-6 Luna vs 1,000,000 for Claude Sonnet 5.5, with Luna's usable input capped at 922,000

• Maximum output - 128,000 tokens on both

• Long-context pricing - Luna reprices the whole request above 272,000 input tokens to $0.20 / $0.75; Sonnet 5.5's card has no equivalent cliff, which is a quiet advantage on 300K-token prompts

• Input modalities - text, image and file on both; neither takes audio or video

• Reasoning control - GPT-6 Luna offers none, low, medium (default), high, xhigh and max, and can be switched off entirely; Claude Sonnet 5.5 has five levels from low to max with high as the default, and reasoning is always on

• Knowledge cutoff - May 18, 2026 for GPT-6 Luna; A​nthropic does not publish a comparable date for Sonnet 5.5

• Release - GPT-6 Luna on September 22, 2026; Claude Sonnet 5.5 on September 28, 2026, six days apart

• AA Intelligence Index v5.0.1 - GPT-6 Luna 28 at max effort vs Claude Sonnet 5.5 42 at its default high effort and 55 at max

• Time to first answer token - GPT-6 Luna 79.33 s vs Claude Sonnet 5.5 10.21 s on the same board

• Cost per Intelligence Index task - GPT-6 Luna $0.15 vs Claude Sonnet 5.5 $1.64

Twenty times the price, eleven times the bill

The list price ratio is a clean 20x. The measured ratio on a real workload is not, because cost per task also counts how many tokens each model spends getting to an answer. On Artificial Analysis's index run, Luna consumed $136 of tokens to complete the suite and Sonnet 5.5 consumed $1,268 - about 9.3x. Luna's per-task figure of $0.15 against Sonnet's $1.64 is a ratio of 10.9.

That difference runs in Luna's favour and against it at the same time, depending on your workload. Luna emits a lot of reasoning tokens for a cheap model: at maximum effort it is the chattier of the two per unit of score, which is why the 20x list ratio compresses to roughly 10x on a bill. On a workload where output dominates - long generations, iterative code edits, anything that writes paragraphs - the compression gets worse for Luna, because its output rate is the expensive half of the card. On a workload where input dominates - retrieval, log reading, document QA over a fixed corpus - the 20x holds almost exactly, and caching makes it better still.

One more arithmetic note that changes the answer for a team already paying A​nthropic prices: Sonnet 5.5 is described by A​nthropic as costing up to 30% less per task than Sonnet 5 despite an unchanged rate card, because it needs fewer tokens to do the same work. A team moving from Sonnet 5 to Sonnet 5.5 gets a real discount without changing a price sheet. A team moving from Sonnet 5.5 to Luna gets a 10x discount and a very different model.

Where the gap is a chasm, and why the harness matters

Terminal-Bench 4.0 is the clearest separation on the board: 43.9% for Sonnet 5.5 against 12.6% for Luna, measured by the same evaluator under the same conditions. Terminal-Bench-Science is worse for Luna still - 34.3% versus 8.6%. If the work is a shell session, a build system, a failing test suite that has to be diagnosed across files, the twentyfold discount is not a discount at all. Luna will not finish the task often enough for the price to matter.

Careful with the numbers you quote for Sonnet here, because there are two of them and only one is comparable. A​nthropic's launch post reports 70.6% on Terminal-Bench 4.0 - more than 26 points above what the independent board measured. Neither figure is dishonest; the vendor number comes from A​nthropic's own harness, tool policy and prompt scaffolding, and Terminal-Bench scores move substantially with all three. Deepnote made the same point from the other direction in its Agent Initiative Benchmark: "the results reflect the models only, not the harness" and a different prompt or tool policy "will move every number here." Treat 70.6% as what Sonnet 5.5 achieves in A​nthropic's setup and 43.9% as what it achieves in someone else's. For comparison shopping, the second is the one to use, because you are not going to run A​nthropic's setup.

The same caution applies to Luna's own vendor numbers. O​penAI reports GPT-6 Luna at maximum effort scoring 66.6% on DeepSWE v1.1, "comparable to Claude Opus 5 and Fable 5 at medium effort," at 93% to 96% less cost per task. That is a vendor-reported result on a vendor-run evaluation, and it says nothing about Terminal-Bench, where the independent number is 12.6%. Both sentences describe the same model.

Where the gap collapses to a rounding error

FrontierCode-AA 2.0 asks whether a coding agent produces a change that is ready to merge into a real repository. Luna scores 25.2. Sonnet 5.5 scores 25.7. Half a point, on a twentyfold price difference, and both far below the leaders on a benchmark where a strong model clears 46%.

Read that carefully before drawing the obvious conclusion. A low absolute score on a hard benchmark means both models fail most of the time; when two models both fail most of the time, a half-point difference is mostly noise. The useful claim is not "Luna codes as well as Sonnet 5.5." It is narrower: on this style of task, at this difficulty, the extra twentyfold buys nothing measurable. If your workload looks like FrontierCode - self-contained edits to a known codebase, reviewed by a human before merge - Luna is a defensible first pass whose output a reviewer was going to rewrite anyway.

Two other measures land in the same neighbourhood: AutomationBench-AA puts Sonnet 5.5 at 59.4% and Luna at 53.2%, and GDP.pdf at 25.2% against 22.8%. Neither is a tie, but both are inside the range where prompt engineering and tool choice can flip the ordering on your own data.

The two places Luna is genuinely worse, not just different

Knowledge and latency, and neither is fixable with a price cut.

AA-Omniscience v1.1 measures whether a model knows what it does not know. Sonnet 5.5 nets +19; Luna nets -8. A negative net score on that evaluation means the model's fabricated answers outnumber its correct ones often enough to drag the total below zero. That is the single most consequential number in this comparison for anyone building an agent that answers without a human in the loop, and it is the number that a benchmark focused on coding will never surface. Combined with Luna's 38.5 on Humanity's Last Exam against Sonnet 5.5's 56.0, the picture is consistent: Luna is a competent instruction-follower with a thin factual floor.

Latency is the second. Sonnet 5.5 reaches its first answer token in about 10 seconds; Luna takes about 79 at maximum effort on the same board. Turn the effort dial down and that gap narrows - Luna's reasoning budget is the cost - but then you are comparing a lower-effort Luna against a full-strength Sonnet, and the score gap widens accordingly. For an interactive product, this is often the deciding factor regardless of price. For a batch job that runs overnight, it is irrelevant.

The one benchmark where they nearly tie on a real job

Artificial Analysis measures models on questions. Deepnote's Agent Initiative Benchmark measures them on a task: fifty simulated workplace situations inside a fake company, where the model reads Slack and decides whether to investigate, where to reply, and whether to reply at all. Claude Sonnet 5.5 scored 6.65 at $0.42 per situation and finished in 15 seconds. GPT-6 Luna scored 6.38 at $0.0070 and took 50 seconds.

Two things follow. The quality difference on this task type - a quarter of a point - is smaller than the interval the benchmark reports, which is why Deepnote groups the two of them with DeepSeek V4.1 Flash, Mistral Large 4 and Claude Haiku 5.5 in a cluster it says "lands within 0.6 points with overlapping intervals." And the cost difference is 60x, not 20x, because Sonnet 5.5 spends roughly twenty thousand tokens per situation at high effort while Luna spends far less.

The split is in discipline rather than capability. Sonnet 5.5 held back on 6 of the 10 situations that called for silence; Luna held back on 4. Sonnet recorded 11 perfect runs and 7 zeros out of 50; Luna recorded 9 and 9. If the job is to be right more often and quiet when unsure, the expensive model earns its money. If the job is to read everything so that a second, stricter gate can decide what to surface, the cheap one is the only version that fits the budget.

The feature that actually separates them is an off switch

GPT-6 Luna accepts reasoning_effort: "none". Claude Sonnet 5.5 does not - reasoning is always on across its five levels, which A​nthropic recalibrated relative to Sonnet 5, so an inherited integration should re-tune its setting rather than carry the old one over.

That single difference explains most of the cost curve. A classification call, a routing decision, an extraction from a known schema: none of these need a reasoning trace, and Luna can be told to skip it. The same work on Sonnet 5.5 pays for reasoning whether it helps or not. It also explains a limit worth knowing before you build around it: on Chat Completions, Luna only supports function calling when reasoning effort is set to none, so anything that needs tools has to go through the Responses API. Sonnet 5.5 takes tools on either surface, through the native Messages API or the O​penAI-compatible format.

O​penAI's Decisions API, generally available since October 22, 2026, is the narrowest expression of the same idea: typed answers - a probability, a choice, a rubric score - from GPT-6 Luna and no other model, at roughly a tenth of the latency O​penAI reports for the same model through Responses. If the question is "does this ticket contradict this order record," Luna can now answer in its native shape without generating prose at all.

Choosing between them

• Long-running terminal work, build repair, multi-file debugging - Claude Sonnet 5.5. Luna's 12.6% on Terminal-Bench 4.0 is not a budget question.

• Anything where a wrong answer is expensive and nobody checks it - Claude Sonnet 5.5, on the strength of the +19 versus −8 Omniscience split.

• Interactive surfaces where 79 seconds to first token is a product defect - Claude Sonnet 5.5.

• High-volume classification, extraction, routing, or a first-pass reader in front of a stricter gate - GPT-6 Luna, at effort none or low, where none of Sonnet's advantages are in play.

• Self-contained code edits that a reviewer will read anyway - GPT-6 Luna as the first pass, Sonnet 5.5 as the escalation path. FrontierCode's half-point difference is what makes the split arguable rather than reckless.

Running both is the option the rate cards do not advertise but the numbers support. A tiered setup - Luna for the 90% of calls that are cheap and narrow, Sonnet 5.5 for the 10% that decide something - costs closer to Luna's bill than to Sonnet's while keeping Sonnet's judgment where judgment matters. OrcaRouter puts both models behind one O​penAI-compatible key alongside more than 200 others, at 0% markup with provider list price passed through, so the 20x ratio on the sheet is the 20x ratio on the invoice and a vendor price change reaches you the same day. Because both accept the O​penAI chat-completions format, the tier split is a routing rule rather than a second integration: the routing DSL sends a request to Luna first and escalates on a condition, automatic failover covers a provider slowdown, and no code changes when a new model lands in the catalogue.

What would change the answer

Luna's current position is unusually dependent on one number moving: its factual floor. A point release that lifts AA-Omniscience from negative to positive would remove the strongest argument for paying twenty times more, because the terminal-work and latency gaps are workload-specific while a model that invents facts is a liability everywhere. Conversely, a Sonnet 5.5 price cut - the kind A​nthropic has done before, and which a 0% markup router would pass through immediately - would narrow the ratio enough to make the tiering question moot for many teams.

Absent either, the honest summary is that these are not substitutes. GPT-6 Luna is a cheap, fast-to-prompt, thin-on-facts reader with a switch that turns thinking off. Claude Sonnet 5.5 is a slower-to-start, expensive, reliable writer that always thinks. The mistake is treating the 20x as a verdict. It is a budget for the parts of the job where the difference does not show up - and FrontierCode's half-point margin is the reminder that some of those parts are bigger than the marketing suggests.