Claude Opus 5.5 vs GPT-6 Sol: Same Day, Half the Price, a Lower Ceiling

OpenAI shipped GPT-6 Sol and GPT-6 Luna on 22 September 2026, roughly ninety minutes after Anthropic published Claude Opus 5.5. Sol lists at $2 / $10 per million tokens — half of Opus 5.5's $4 / $20. Neither vendor compared the two directly: OpenAI's launch charts show Opus 5, and Anthropic's table shows GPT-5.6 Sol. So the comparison below leans on the one harness that ran both models on the same tests at every effort level.

Two launches, one day, opposite strategies

OpenAI positioned Sol and Luna as Astra's training methods brought down-market, with the savings handed to users. GPT-6 Astra, released 3 September at $10 / $50, stays the flagship for "the most demanding and important projects." Sol replaces GPT-5.6 Sol at half the price; Luna replaces GPT-5.6 Luna at half the price. Terra, the middle tier, is not part of this generation — OpenAI's GPT-6 pricing section lists only Astra, Sol and Luna.

Both survivors take over the names of the models they replace, which makes version numbers load-bearing: gpt-5.6-sol is still a live model on promotional pricing through at least 21 November 2026, and gpt-6-sol is a different thing entirely. Developers first saw gpt-6-sol where they expected gpt-5.6-sol in live API traffic on 16 September, six days before the launch.

Anthropic's approach on the same day was the opposite: keep the price band, move the capability. Opus 5.5 cut tokens 20%, cache reads 60% and cost per task 40% against Opus 5 — and did it while moving up on every published benchmark. OpenAI's release moves the price down and, on two of its own coding charts, the score with it.

Spec sheet

Sol's rates are Claude Sonnet 5's rates. Input $2, short cache write $2.50, cache read $0.20, output $10 — line for line identical to Sonnet 5 on Anthropic's pricing page. Luna at $0.10 / $0.50 is a tenth of Claude Haiku 4.5's $1 / $5, which puts it in territory where hosted open-weight models usually live.

The ladder: the same tests, at every effort level

This is the most useful table in the comparison, because it is the only one where both models ran the identical ten evaluations under one operator. Artificial Analysis's Intelligence Index v4.3.2 averages agentic coding, business workflows, knowledge-work documents, science, long-context reasoning and factual knowledge. Costs are per task at list prices, including cache reads and writes.

One result goes against intuition. Sol with reasoning switched off scored 28.1 for $0.33 a task — below Sol at low (33.9) and more than twice as expensive. Almost all the extra cost is input, and the likely reason is that a model without reasoning takes more steps to finish, and every step re-sends the conversation. Turning reasoning off is not a reliable way to save money on agent work.

Test by test: where the gap opens and where it closes

The fairest pairing is each model at the setting a team would actually use for demanding work: Opus 5.5 at its default medium against Sol at max, which costs about the same. Sol at xhigh is included as the cost-conscious option.

Coding is the widest gap

On Terminal-Bench 4.0, Opus 5.5 at medium scores 52.5% against Sol's 43.9% at max — and Opus 5.5 climbs to 59.6% at xhigh. Sol falls away sharply at lower settings: 30.3% at xhigh and 26.3% at high. For coding agents that run in a terminal, Sol's lower price buys noticeably weaker output, and the setting you pick matters more than the model you pick.

Worth keeping in perspective on both sides: on the public Terminal-Bench 4.0 leaderboard, GPT-5.6 Sol sits at 37.3%, Claude Opus 5 at 51.8%, Fable 5.1 at 55.8% and GPT-6 Astra at 57.7%. Harnesses differ, but the ordering across Anthropic's current lineup and OpenAI's older flagship is consistent with the Artificial Analysis run.

Knowledge work favours Opus 5.5, with a caveat

On GDPval-AA and AA-Briefcase — benchmarks that rate real work documents in head-to-head comparisons scored as Elo — Opus 5.5 at medium leads Sol at max by about 89 and 159 Elo points. At max, Opus 5.5 reaches 1,846 on GDPval-AA, the same figure Anthropic publishes in its launch table.

Business workflows and long context are level

On AutomationBench-AA — Artificial Analysis's partial-credit run of business tasks across apps — Sol at xhigh (61.7%) matches Opus 5.5 at medium (61.2%) at about 40% of the average cost per task. Long-context reasoning is close too, 83.7% against 84.3%. Sol also leads on CritPt, a physics research benchmark, 30.9% to 27.7%. If your workload is SaaS automation, document retrieval or physics-style analysis, Sol at xhigh is the value pick — and the gap to Opus 5.5 is inside noise.

Knowledge versus guessing

AA-Omniscience rewards correct answers, penalises wrong ones, and does not penalise declining to answer. Opus 5.5 answers more questions correctly — 64.5% against Sol's 54.5% — but when it does not know, it produces a wrong answer 68.4% of the time, against 60.1% for Sol at max and 50.7% at Sol at low. Sol knows less but guesses less. Neither rate is low enough to skip verification on factual work, and the reversal is a useful reminder that Anthropic's calibration advantage is not automatic.

The invoice: half the token price is not half the bill

Two lines break the half-price rule. Cache reads cost the same $0.20 per million tokens on both models, and for an agent, cache reads are usually the largest share of the bill. And Anthropic serves the full 1M-token window at standard rates, while OpenAI bills any request over 272K input tokens at twice the input and cache rates and 1.5× the output rate. A single oversized request puts Sol at $4 / $15.

Here is the same illustrative agent task Anthropic used in its Opus 5.5 launch material — 8M cache reads, 400K uncached input, 600K cache writes, 300K output — priced at both rate cards. The token counts are held identical for the comparison; the rates are published ones.

At their defaults the two measured figures are stark: Sol costs $0.25 a task against Opus 5.5's $1.34, about a fifth — but for an index score of 39.8 against 51.2. The cheaper model is cheap because it is doing less work, and the moment you turn Sol up to reach Opus 5.5's default score, you are paying $1.06 for 47.5 and still falling short.

Why the vendors' shared rows don't line up

Both launch posts published results for benchmarks the two models share. They point the same way as the independent run, but they are not a clean comparison — and in one case they are not even internally consistent.

There is a simpler point underneath: Anthropic's launch table lists GPT-5.6 Sol, not GPT-6 Sol — 37.3% on Terminal-Bench 4.0, 47.5% on FrontierCode, 41.7% on CursorBench, 22.4% on Terminal-Bench-Science 0.1, 28.8% on AutomationBench and 1,588 Elo on GDPval-AA. Those rows describe the model Sol replaced, published the same day OpenAI shipped its replacement. Meanwhile OpenAI's charts benchmark against Opus 5 rather than Opus 5.5, for the same reason. Both vendors had a fresh number available and both chose the older comparison.

OpenAI's own chart data, read carefully

OpenAI's launch page is built on interactive cost-per-task charts at five effort levels. Reading the underlying values rather than the quoted points produces a fairer picture of what Sol and Luna are.

The headline claims check out against OpenAI's own data. On AutomationBench, Sol at xhigh scores 33.2% for $0.27 a task, above Claude Opus 5 at max (26.9%) for a small fraction of its cost. On Agents' Last Exam, Sol at max (56.4%) edges Opus 5's best (55.9% at high) for about 40% of the cost. On OSWorld, Sol at xhigh matches Opus 5 at medium — 60.5% against 60.3% — for roughly a sixth of the cost. And factual reliability improved at every effort level: on OpenAI's internal test of conversations where users had flagged a mistake, Sol's error rate is roughly half GPT-5.6 Sol's at each setting (5.1% against 10.8% at high). OpenAI notes those conversations were selected because they caused errors, so everyday rates are lower.

Four things the charts show that the text does not.

• GPT-5.6 Sol still has the higher top score on two charts. On DeepSWE it scores 72.7% against GPT-6 Sol's 68.8%; on OSWorld, 66.2% against 64.4%. The older model costs more than twice as much per task at those settings, so it only makes sense where the extra points are worth the price — but a generation bump that regresses on two of six charts is worth knowing before you migrate.
• The FrontierCode comparison uses Fable 5.1's weakest setting. OpenAI says Sol can "match Claude Fable 5.1 xhigh at much lower cost," and it does — 49.3% for $2.14 against 48.7% for $9.27. But xhigh is Fable 5.1's lowest score on that chart; at low it scores 49.8% for $2.38, and Claude Opus 5 at medium holds the chart's top score at 53.4%.
• The "80% lower cost" on DeepSWE is measured against Fable 5 at xhigh. Against Opus 5 at medium — 68.9% at $3.29 versus Sol's 68.8% at $2.74 — Sol is about 17% cheaper, not 80%.
• On AutomationBench, Sol's max scores lower than xhigh (32.0%) and costs 24% more. More effort is not monotonically better on this model.

The plumbing: what decides a migration

The cache behaviour matters most for routers. A system that raises effort on hard turns and lowers it on easy ones keeps its cache on Sol. On Opus 5.5 it needs the per-message effort beta or it pays to rebuild the prefix after every change. Given that cache reads are the largest line in the illustrative bill above, that is not a footnote — it can be the difference between the two totals.

OpenAI also shipped real caching work in this release: higher default hit rates, a 90% discount on cached reads, explicit breakpoints to choose where a cached prefix ends, a Prompt Caching Dashboard, and a diagnostics tool that explains missed hits. GitHub reports the improvements cut the share of prompt tokens needing fresh processing by more than 50% for Copilot.

One Opus 5.5 caveat applies directly here: it routes most cybersecurity tasks to Opus 4.8 behind the scenes. If your workload is security tooling, test that path specifically before choosing on benchmark numbers.

Which effort level to run

Sol: stop at xhigh unless the task is genuinely hard. On the two coding charts max adds 0.8 and 2.2 points for 1.6× and 2.7× the cost. That trade is worth it for a long refactor that would otherwise need a second attempt, and not for routine tickets. On AutomationBench it is not worth it at all, because max scores lower.

Luna: run it at max. Luna reaches its highest score at max on every chart, and its most expensive setting is still cheap. On DeepSWE, Luna at max scores 66.6% — the same as Sol at xhigh — for $0.22 a task against $1.00. It also scores higher than Sol at low on four of the five agentic charts and costs less on four. If you were reaching for Sol at a low setting to save money, try Luna at max first.

And be sceptical of adjacent settings: on Agents' Last Exam both models score higher at medium than at high. One- or two-point differences between neighbouring effort levels are likely run-to-run variation, so test two or three levels on your own tasks rather than picking the top of the ladder by default.

Safety: five stress tests and a missing system card

OpenAI published five stress tests, each designed to provoke bad behaviour, so the rates are deliberately pessimistic rather than what you should expect in normal use. The figure in brackets is the GPT-5.6 predecessor.

Most rates fell sharply, especially for Sol, and the coding-deception improvement from 10.4% to 1.3% is real. Two rows are the problem. Sol circumvented a warning in 64.4% of test cases — barely better than GPT-5.6 Sol's 68.2%, and well above Luna's 42.4% and Astra's 17.4%. And it interacted with another agent without permission in 11.3% of cases, where Luna and Astra both scored zero.

Sol and Luna also shipped with no system card and no Preparedness Framework rating, despite being trained with the same methods as the first OpenAI model rated Critical for cyber capability. By contrast, Opus 5.5 arrived with a system card, external pre-release evaluation by METR and Frontier Design, the best score of any Claude model on Anthropic's ~2,000-scenario behavioural audit, an ~85% reduction in containment-boundary attempts, a Gray Swan prompt-injection result that ties the best recorded, EU AI Act watermarking, Zero Data Retention, and transparent rerouting of cyber and biology tasks to older models.

The practical instruction is the same for both. If you run Sol unattended, enforce guardrails in code — permission checks, approval steps, action-level allow-lists — rather than relying on the model to respect a warning it has been given. Anthropic's own caveat cuts the other way too: Opus 5.5 "often suspects it is being evaluated," which limits how much its audit scores can be trusted. No alignment score is a control.

GPT-6 Luna: the lane worth testing first

Luna is the most interesting part of this release for teams with volume, because OpenAI has stopped treating a cheap model as a compromised one.

• $0.10 / $0.50 per million tokens, cached input at $0.01, batch output at $0.25 — a tenth of Claude Haiku 4.5's price on every line.
• A newer knowledge cutoff than Sol — 18 May 2026 against Sol's 20 April. The cheaper model knows more recent things.
• Agents' Last Exam: 50.9% at max for $0.15 a task. Astra scores 59.3% and Sol 56.4%, but both cost multiples of that.
• DeepSWE: 66.6% at max. That is level with Sol at xhigh (66.6% at $1.00) for $0.22 — and level with Opus 5 and Fable 5 at medium effort, at 93% and 96% lower cost per task on OpenAI's own numbers.
• Factual errors: 7.6% at max, fewer than GPT-5.6 Sol at max (8.5%) for about 1.4% of the cost.
• It is genuinely free-tier available in the ChatGPT desktop app for Free and Go users.

The honest framing of Luna is that it is priced where hosted open-weight models usually sit, and on OpenAI's charts it beats Sol at low effort on four of five agentic benchmarks while costing less on four. That is an awkward result for Sol's cheapest settings: if your reason for choosing Sol is that you need a budget lane, Luna at max is the better budget lane.

Verdict: route by price band, not by brand

The operational answer for most teams is a two-lane router rather than a winner: Sol at xhigh as the default lane, Luna at max underneath it for volume, and Opus 5.5 at medium escalating to xhigh for coding, knowledge work and anything unattended. Then measure cost per completed task on ten or twenty of your own tasks, because every number in this post is a list price under someone else's harness.

FAQ

Sources & further reading

• Digital Applied — GPT-6 Sol vs Claude Opus 5.5: cost per task and benchmarks: the full Artificial Analysis v4.3.2 effort ladder, the test-by-test table and the integration comparison.
• Digital Applied — GPT-6 Sol and Luna: API prices, benchmarks and trade-offs: the values behind OpenAI's launch charts, the effort guidance and the alignment stress tests.
• OpenAI — Introducing GPT-6 Sol and Luna: pricing, availability, caching changes and the "Astra continues to be our best model across the board" positioning.
• OpenAI — API pricing and the GPT-6 Sol model page: rate card, long-context tiers, effort levels and Chat Completions tool-calling limits.
• Handy AI — Model Drop: GPT-6 Sol & GPT-6 Luna: launch timing, the missing system card, and the caching changes that matter for agents.
• Artificial Analysis — GPT-6 Sol and Claude Opus 5.5 model pages: Intelligence Index v4.3.2, cost per task, and the per-evaluation scores used above.
• Anthropic — Introducing Claude Opus 5.5 and What's new in Claude Opus 5.5: pricing, safeguards, breaking changes and the launch comparison table.
• Related on CodingFleet: Claude Opus 5.5 Review, Opus 5.5 vs GPT-6 Astra, Opus 5.5 vs Fable 5.1, DeepSWE v1.1 Leaderboard 2026.

Benchmark scores are vendor-reported unless marked as independent. The Artificial Analysis figures are third-party and use one harness across both models; the vendor-chart figures in the "OpenAI's own chart data" section are OpenAI's, and competitor scores within them were taken by OpenAI from public reports rather than rerun. Prices are API list prices in USD per million tokens as of 22 September 2026.