Verdict first: route everyday agent work to GPT-6 Luna at $0.10 input / $0.50 output per million tokens, step up to GPT-6 Sol at $2 / $10 when the task needs near-flagship depth, and pick Claude Opus 5.5 at $4 / $20 when long supervised coding runs justify the premium. All three launched on September 22, 2026 within hours of each other — duelling cheaper models against a backdrop of safety debate — and each vendor claims its cheaper tier now does work that used to require the flagship. Everything below exists to defend that verdict: what cheaper-beats-flagship means, a head-to-head grid in which every figure traces to a cited source, per-model best-fit notes, the self-host question, and what to verify on your own workload. This comparison reuses the pilot format from our decision-models comparison and extends the GPT-6 Astra launch analysis, whose $10 / $50 pricing is the flagship baseline here.
When cheaper beats flagship
Cheaper beats flagship when the task fits inside the smaller model's strengths and the savings compound across volume. OpenAI's claim is concrete: Sol at xhigh effort scores 33.2% on AutomationBench at $0.27 per task, versus 26.9% for Claude Opus 5 at max effort at 11.1 times the cost, and Luna at max effort exceeds GPT-5.6 Sol at medium effort at one tenth of its cost on OSWorld. Anthropic's mirror claim: Opus 5.5 at default effort beats Opus 5 at max effort for about a fifth of the cost on Terminal-Bench and beats GPT-6 Astra's top FrontierCode score for about a fifth of the cost per task. The pattern is the same one behind scoring beating sampling: when the workload is bounded and repeated, cost per completed task is the metric that matters, not headline capability. Flagships keep the work that is genuinely open-ended, unrepeatable, or safety-gated — everything else should face a per-task cost comparison first. Independent first impressions, including Simon Willison's price-war roundup, already show practitioners switching defaults to Sol and Opus 5.5 on exactly this logic.
Head-to-head: Sol vs Luna vs Opus 5.5
The grid below summarizes each vendor's published position at the time of writing. Figures marked vendor-reported come from the vendor's own announcement; anything the announcements do not state is marked unverified rather than guessed.
| Criterion | GPT-6 Sol (OpenAI) | GPT-6 Luna (OpenAI) | Claude Opus 5.5 (Anthropic) |
|---|---|---|---|
| Price per 1M tokens | $2 input / $10 output, vendor-reported 50% cut vs GPT-5.6 Sol promo (announcement; pricing page) | $0.10 input / $0.50 output, vendor-reported 50% cut vs GPT-5.6 Luna promo (announcement; pricing page) | $4 input / $20 output, vendor-reported 20% cut vs Opus 5 (announcement; pricing page) |
| Cached input | 90% discount on cached reads with improved hit rates, vendor-reported (announcement) | 90% discount on cached reads with improved hit rates, vendor-reported (announcement) | $0.20 per 1M cache reads, vendor-reported 60% cut vs Opus 5 (announcement) |
| Speed claim | No vendor tokens-per-second figure; faster in cost-per-task terms only (unverified as raw speed) | No vendor tokens-per-second figure; faster in cost-per-task terms only (unverified as raw speed) | Vendor-reported 30%+ faster output than Opus 5; Fast mode up to 2.5x at $8 / $40 (announcement) |
| Context window | Unverified — the launch announcement states no context window; check the model docs before building | Unverified — the launch announcement states no context window; check the model docs before building | Unverified — the launch announcement states no context window; check the model docs before building |
| Signature result | AutomationBench 33.2% at $0.27 per task, vendor-reported (announcement) | DeepSWE 66.6% at max effort, comparable to Opus 5 at medium for 93% less per task, vendor-reported (announcement) | Terminal-Bench 66.4% and OSWorld 81.8% partial, vendor-reported (announcement) |
| Best fit | Heavy everyday professional and coding work at flagship-adjacent quality | High-volume agent loops, retrieval, and iteration where unit cost dominates | Long supervised coding migrations and audits where fewer steps win |
GPT-6 Sol: best for heavy work at near-flagship depth
Sol is the price war's middleweight: $2 input / $10 output per million tokens, half of GPT-5.6 Sol's promotional pricing, with 90% cached-input discounts and higher default hit rates. The vendor's per-task figures are the argument: 56.4% on Agents' Last Exam at 60% lower cost per task than Opus 5's best, FrontierCode matching Fable 5.1 at much lower cost, and OSWorld 60.5% against Opus 5's 60.3% at roughly 80% lower cost. Factuality improves too — about half as many mistakes as its predecessor on OpenAI's internal error-flagged conversations. Independent observers moved fast: Willison made Sol a default in Codex within a day. Best for: professional workflows, coding agents, and computer-use tasks where Astra's $10 / $50 rate (our Astra analysis) cannot clear the budget. Unverified: raw throughput and context window — the announcement gives neither, so confirm both in the docs before promising latency to anyone.
GPT-6 Luna: best for high-volume loops where unit cost dominates
Luna is the cheapest capable model either frontier lab has shipped in this cycle: $0.10 input / $0.50 output per million tokens, which Willison notes is undercut among OpenAI models only by far weaker Nanos. The vendor positions it as an iteration engine — 5.4 points up on its predecessor on AutomationBench at 58% lower cost per task — and early independent use agrees: Willison upgraded a Datasette agent demo to Luna for SQL and app-building work. Note the pricing mechanics: with 90% cached-input discounts, long agentic conversations where most input tokens are cached get cheaper still, which is exactly the shape of retrieval and re-check loops. Compare unit economics on the prices page before routing. Best for: retrieval grounding, re-ranking, test-and-retry loops, and every background job where you currently hesitate to call a flagship. Unverified: throughput and context window, as with Sol.
Claude Opus 5.5: best for long supervised coding runs
Opus 5.5 answers a different question: what if the expensive model gets dramatically more efficient instead of dramatically cheaper. Input and output tokens fall 20% to $4 / $20, cache reads fall 60% to $0.20, and typical workloads cost 40% less than Opus 5, while output arrives 30%+ faster. The vendor's efficiency stories are vivid — a 200,000-line audit finished in under three hours where Opus 5 took over twenty on 2.5x the tokens, and a translation job finished in 9.5 hours versus 12 for Fable 5.1 at 51% lower cost — backed by benchmark leads in Terminal-Bench (66.4%), FrontierCode (54.4%), OSWorld (81.8% partial), and Humanity's Last Exam with tools (67.7%). Two cautions come from the vendors and observers themselves: Anthropic says Opus 5.5 often suspects it is being evaluated, which complicates predicting real-world behavior, and Willison found max effort over-thinking to the point of hitting the 128,000-token output cap on a trivial prompt. Best for: codebase migrations, audits, and supervised long-horizon work where fewer, better steps beat cheaper tokens. Check score context on the benchmarks page.
Self-host vs API: none of the three ships weights
Every contender here is API-only, so the self-host question resolves to which separate open model you pair them with — the same split behind the decision-models comparison.
| Dimension | GPT-6 Sol / Luna | Claude Opus 5.5 |
|---|---|---|
| Access | Managed API as gpt-6-sol and gpt-6-luna; ChatGPT Work and Codex for paid plans, Luna also on desktop for Free and Go users, not yet in Chat (announcement) | Managed API as claude-opus-5-5 on all major clouds; zero-data-retention option like prior Opus models (announcement) |
| Weights | None published; self-hosting means a different open model (unverified — no weights statement either way beyond API-only availability) | None published; self-hosting means a different open model, plus preserved-thinking anti-distillation limits on context editing (announcement) |
| Safety gating | Builds on Astra alignment work with lower misleading-claim rates on coding, vendor-reported (announcement) | Fable-class safeguards on cyber, biology, and distillation; risky requests reroute to an older model, and broader safety context is covered in the September 23 report |
| Self-host pairing | Keep drafts, evals, and private data on a local model; call Sol or Luna for the passes that need frontier depth | Same hybrid pattern: local first, Opus 5.5 for the supervised runs where its per-task efficiency wins |
Verify on your workload
Three caveats keep this comparison honest. First, every per-task cost figure is vendor-reported on vendor-chosen harnesses and effort levels — OpenAI's tables compare research-environment scores against competitor figures from public reports, and Anthropic notes benchmark margins at these capability levels are a less reliable guide than real-world use. Re-run the head-to-head on your own traffic before quoting savings. Second, vendor benchmarks omit failure modes you will meet in production: safeguard reroutes that hand work to older models, paused tasks awaiting review, and effort settings that can over-think simple prompts into output-cap failures. Third, prices move: GPT-5.6 faces a scheduled 25% increase in November against which GPT-6 is half price, and anything numeric here describes a cited claim at publication time, not a standing quote — re-check prices and the vendor pages before budgeting.
How we compared
Scope honesty: we compared published announcements and independent reporting, not models on our own hardware. No synthetic benchmark scores, no invented win rates, no leaderboard. Every figure links the source it came from, vendor-reported numbers are labeled as such, and anything the sources do not state is marked unverified.
Our method follows the comparison pilot: verdict first, then evidence, then method, then questions. We read the launch announcements (OpenAI on Sol and Luna; Anthropic on Opus 5.5), the independent price analysis (Willison's roundup), the wider news context (the September 23 report), and the vendor pricing pages (OpenAI; Anthropic), cross-checked against the Astra flagship baseline and the prices page. Readers should reproduce the comparison rather than trust our adjectives: each linked source contains enough detail to re-check every row of the grid in an afternoon.
Questions and answers
When does cheaper beat the flagship?
When the workload is bounded, repeated, and measurable per completed task: agent loops, retrieval, regression-capable coding, and batch professional work. Route those to Luna or Sol first and keep Astra or Fable for open-ended, unrepeatable, or safety-gated work — then compare cost per finished task, not cost per token.
Which cheap flagship should I try first?
Luna if unit cost dominates and the task fits a smaller model; Sol if you need near-flagship depth at roughly flagship-adjacent quality for a fifth of Astra's rate; Opus 5.5 if long supervised coding runs where fewer steps and faster output outweigh the higher token price. All three are API-only, so budget a fallback and log every run.
Can I self-host any of them?
No — none of the three ships weights. Self-hosting means pairing a different open model for private data, drafts, and evals, and calling the managed API only for the passes that need frontier depth. Size your hardware first and confirm licences before promising anything local.
What should I distrust in this space right now?
Unreproduced vendor tables, effort-level comparisons that flatter the house model, and prices quoted without dates. Context windows, throughput figures, and anything the announcements omit are unverified until the docs or an independent reproduction confirm them. Re-check the vendor pricing pages and rerun the comparison on your own workload before budgeting.
Where do I go next on this site?
Read the flagship baseline in the Astra launch analysis, compare unit economics on the prices page and score context on the benchmarks page, revisit the format pilot in the decision-models comparison, then watch the safety story behind these launches unfold in future digests.