Claude Mythos 5対GPT-6.1 Sol：一方には測定済みのコスト曲線があり、他方には空白の行がある

Artificial Analysis prices every Intelligence Index run it publishes, and the row for GPT-6.1 Sol reads $0.72 per task. There is no equivalent row for C​laude M​ythos 5, and there will not be one while the model stays behind the verification wall, because an evaluator cannot score a model it cannot call. That asymmetry is the whole comparison, and it outranks every benchmark number either vendor has printed. GPT-6.1 Sol is O​penAI's near-flagship from September 29, 2026, at $2.00 per million input tokens and $10.00 per million output, scoring 51.8 on the Index at maximum reasoning effort for $0.72 a run. C​laude M​ythos 5 is A​nthropic's restricted cyber and biology flagship from June 9, 2026, priced at the M​ythos tier's $10.00 / $50.00, refreshed as C​laude M​ythos 5.1 on September 1, and handed out through an application rather than a rate card. One of these has a measured cost-of-intelligence curve. The other has a price and no curve, and after October 6 the second one also has a form to fill in.

9週間と1つの扉を隔てた2つのリリース

C​laude M​ythos 5 is the restricted member of A​nthropic's June 9 frontier generation. A​nthropic ships each frontier step twice from the same underlying weights: a public Claude Fable SKU with the full safeguard stack, and a M​ythos SKU with the cyber and biology safeguards tuned down for vetted defenders and researchers. C​laude M​ythos 5 is the June restricted SKU, Claude Fable 5 is its public twin, and C​laude M​ythos 5.1 and Claude Fable 5.1 repeat the pattern in the September 1 refresh. The model is ASL-3 classified — the tier A​nthropic reserves for capabilities it judges could meaningfully uplift cyberoffense — and its reputation was built on red-team results rather than a published benchmark table: an undiscovered OpenBSD bug roughly 27 years old, an FFmpeg vulnerability that had survived about 16 years and five million automated scan hits, a browser exploit chaining four vulnerabilities to escape both renderer and OS sandboxes, and a Linux kernel privilege escalation.

GPT-6.1 Sol arrived nine weeks later with the opposite disclosure posture. It is positioned one tier below GPT-6 Astra in O​penAI's G​PT-6 series, replacing GPT-6 Sol at the same headline rates with cached input halved to $0.10 per million. It reached general availability on September 29, 2026 — the same day as GPT-6 Sol and G​PT-6 Luna — in ChatGPT Work and Codex on Plus, Pro, Business, Enterprise and Edu; it is deliberately absent from the consumer chat product and from Free and Go, and on Enterprise and Edu an administrator has to switch it on. The snapshot list holds a single identifier, gpt-6.1-sol, with no dated variant, so the model on the page and the model you call are the same thing.

両方の形状を、次元ごとに：

• Context and output — 1M tokens with 128K maximum output for C​laude M​ythos 5 vs 1,050,000 tokens with 128K maximum output for GPT-6.1 Sol.

• モダリティ — テキスト、画像、ファイルの入力に対応し、どちらもテキスト出力。

• Reasoning control — adaptive thinking that cannot be switched off on the A​nthropic side vs reasoning always on across five effort levels from low to max, defaulting to medium, on the O​penAI side.

• Sampling — C​laude M​ythos 5 accepts the usual A​nthropic parameters; GPT-6.1 Sol does not accept temperature or top_p at all, and tunes depth through effort instead.

• Knowledge cutoff — June 2026 for C​laude M​ythos 5; O​penAI publishes April 30, 2026 for GPT-6.1 Sol.

• Access — an application through the Cyber Verification Program or the Life Sciences Verification Program, with a 30-day data-retention requirement, vs a self-serve API key, plus AWS, G​oogle Cloud and M​icrosoft Azure.

ステッカーが隠している階層を含む料金表

• List price — $10.00 input / $50.00 output per million tokens for C​laude M​ythos 5 vs $2.00 / $10.00 for GPT-6.1 Sol, a fivefold gap on both directions.

• Cache reads — $1.00 per million on the June M​ythos generation, reduced to $0.25 in the September refresh, vs $0.10 per million on GPT-6.1 Sol.

• The long-context cliff — GPT-6.1 Sol holds $2/$10 up to 272,000 input tokens and then moves to $4.00 / $15.00, with the cached rate doubling to $0.20 as well. No equivalent published tier exists for C​laude M​ythos 5.

• Batch — A​nthropic publishes batch pricing at half the M​ythos list rate, $5.00 / $25.00; O​penAI's batch discount is the usual half-rate arrangement.

• Region multiplier — A​nthropic applies a 1.1× multiplier for US-only inference on the M​ythos track.

The tier row is the one most readers skip and it is the one that decides large-prompt economics. This comparison is a 1M-context against a 1.05M-context model, which means both are sold for exactly the workloads that cross into the second pricing tier. On a 400,000-token prompt, GPT-6.1 Sol is not a $2/$10 model — it is a $4/$15 model, and the fivefold gap to the M​ythos rate has narrowed to 2.5× on input and 3.3× on output. At that prompt size on reasoning-heavy work, the two rate cards are closer than either vendor's headline suggests.

各モデルが知られている番号と、それを公開した人

GPT-6.1 Sol's Independent score is Artificial Analysis's, and it arrived with a completeness that is genuinely unusual for a DevDay release. The version running on September 30, 2026 was the evaluator's v4.3.2 Index — a composite of ten evaluations including Terminal-Bench 4.0, GDPval-AA v2.1, AutomationBench-AA, SciCode, Humanity's Last Exam and GDP.pdf — and all three O​penAI models in this piece sit on the same revision, which makes the comparison below like-for-like in a way no vendor's launch table ever is.

• インテリジェンス指数 — 最大エフォート時のGPT-6.1 Solは51.8、それに対してGPT-6 Astraは52.7、GPT-6 Solは47.5。

• インデックスタスクあたりのコスト — GPT-6.1 Sol は $0.72、Astra は $3.26、GPT-6 Sol は $1.05。

• この実行で消費された出力トークン — GPT-6.1 Solが6,720万、Astraが6,000万、GPT-6 Solが7,680万。

最初の2行は一緒に読んでください。というのも、その組み合わせこそが役立つ部分だからです。GPT-6.1 Sol から GPT-6 Astra に移ると、4.5倍の金額で買えるのは Index ポイント0.84分です。この世代の天井がどの高さにあっても、それは一段下のティアにはない形で高くつきます。

Now the A​nthropic side. A​nthropic does not publish a benchmark table for C​laude M​ythos 5 at all. The five-month-old model's numbers travel as figures from its launch coverage, and the two most-quoted ones have provenance problems worth stating plainly:

• ExploitBench, 78.0% — reported by A​nthropic at the June launch, roughly double the 40% it cited for C​laude O​pus 4.8. Vendor-reported, never independently reproduced.

• CyberGym Level 1, 83.8% — the figure most often printed next to the name, and it does not originate with A​nthropic. It appears in a competitor's comparison material, which has its own incentive in how the race is framed.

• Terminal-Bench 4.0, 42.0% for the June generation and 60.9% for C​laude M​ythos 5.1 — A​nthropic's own figures from the September refresh, vendor-reported, and the only agentic-coding number that exists for this track.

So the honest summary of the evidence is lopsided in a specific way. For GPT-6.1 Sol you have a third party's composite score plus a third party's measurement of what the evaluation cost to run. For C​laude M​ythos 5 you have one vendor's security-specific number, one competitor's security-specific number, and an agentic coding score from the vendor. Those sets cannot be subtracted

互いに異なっており、それらを一列に並べるページは比較をでっち上げている。

The gap gets a little less lopsided as of October 15, when A​nthropic began granting a METR team employee-like access — company hardware, office space, internal systems, and the right to publish analyses that A​nthropic cannot veto, only redact in predefined categories. That is the structural fix for exactly the problem this matchup exposes, and it is worth watching whether a restricted model ever acquires an independent row as a result. As of today it has not.

10月6日の変更が実際に変えたもの

On October 6, 2026, A​nthropic rebuilt the Cyber Verification Program into three application-based tiers, folded Project Glasswing into the top one, and turned C​laude M​ythos 5.1 from a partner-only allocation into an access grant any qualifying security organisation can apply for. C​laude O​pus 5.5 and C​laude S​onnet 5.5 were added to every tier at the same time. Defense Access applications are answered in a few days per A​nthropic, Red Team Access reviews take a few weeks, and M​ythos access itself trails approval by roughly five business days. All vendor-stated and independently unconfirmed, and the bar is still a verification standard rather th

ある価格。

What changed is not the price and not the model. What changed is that the door now has a handle. Before October 6 the M​ythos track went to Project Glasswing partners — roughly a hundred vetted organisations — and prior M​ythos Preview holders, with no sign-up flow at all; the June 9 generation had been suspended worldwide two days after launch by a US Commerce Department order, with access restored only for a limited set of vetted US organisations around the start of July. An application form that any qualifying security team can submit is a genuinely different object from a partner allocation, and it is the first time in this model's history that the answer to "how do I get it" has not been "you don't."

オフェンシブセキュリティ、生物学、化学を仕事としていない読者にとって、それは依然として開かれた扉ではなく、閉ざされた扉だ。申請は存在する。適格性は存在しない。

自分のトラフィックでこの比較を実行する方法

ほとんどのチームにとって実用的な問いは、どちらのモデルのほうが優れているかではない。なぜなら、一方のモデルはほとんどのチームには購入できないからだ。実用的な問いは、自分たちが対象条件を満たすかどうかを知るのを待つ間、ルートに何を入れておくかである。

GPT-6.1 Sol sits in the OrcaRouter catalogue at O​penAI's own list price — $2.00 input and $10.00 output per million, $0.10 cached input, with the 272,000-token tier boundary carried through exactly — passed at 0% markup, so a vendor rate change is live here the day it is posted rather than at your next contract renewal. The reason that matters for this particular model is its effort ladder: five settings from low through max, defaulting to medium, and the measured cost of a full Index run ranges from $0.72 at max effort to materially less at lower settings. Tuning effort is the cheapest lever available on this model and the easiest one to get wrong under a fixed timeout, which is where automatic failover across providers earns its place: a route that degrades, or a backend that is simply slow tonight, hands the request to another provider instead of returning an error to your user.

If what you actually wanted from C​laude M​ythos 5 was the capability rather than the reduced safeguard level, Claude Fable 5.1 is the same underlying weights with the safeguards left on, and it is served on the same catalogue at A​nthropic's list price of $10.00 input / $50.00 output with the September refresh's $0.25 cache read in place. Swapping between it and GPT-6.1 Sol is a model-string change, not a second integration. A gated model cannot be routed to the public — eligibility is A​nthropic's decision and no platform changes it — but its ungated twin can, and that is the honest version of this recommendation.

So the decision rule is short. If your work is authorised offensive validation or vulnerability research against systems you own, the October 6 program expansion is the first genuine opportunity this model has offered you, and the 5× premium is the price of a capability tier rather than a benchmark score. For everyone else, GPT-6.1 Sol is cheaper by a factor of five on the sticker, cheaper again on cache, independently measured at a documented cost per task, and callable this afternoon. The blank row on the other side is not a mark against C​laude M​ythos 5 — it is a statement that nobody outside A​nthropic has been able to price what it does, and you should not pay a 5× premium to be the first.

そうした話は決して仮定ではない。このようなリリースの背景にある、記録に残されたAIの障害やエージェントのインシデントはOrcaRouter AIインシデントアーカイブに集められており、各項目には独自の情報源、深刻度、信頼度グレード、係争中フラグが付されています。