Claude Mythos 5 vs GPT-6 Sol: A $50 Output Price Against a $10 One

Claude Mythos 5 costs $50 per million output tokens and cannot be bought without an application. GPT-6 Sol costs $10 per million output tokens and has been sitting in the OrcaRouter catalogue since OpenAI shipped it on 22 September 2026. That is the whole comparison in one line, and it is why most teams that ask the question end up somewhere other than the obvious answer. Claude Mythos 5, released on 9 June 2026, is the specialist: a Mythos-class model with cyber safeguards lifted, sold to vetted defence organisations. GPT-6 Sol is the generalist that OpenAI priced to win volume. The interesting question is not which is smarter. It is what the extra $40 buys, and whether you are allowed to buy it.

The rate card, line by line

• Input — Claude Mythos 5 $10 per million; GPT-6 Sol $2 per million.

• Output — Claude Mythos 5 $50 per million; GPT-6 Sol $10 per million.

• Cached input — Claude Mythos 5 $1.00 per million; GPT-6 Sol $0.20 per million.

• Context — Claude Mythos 5 1M tokens; GPT-6 Sol 1.05M tokens.

• Max output — Claude Mythos 5 128K tokens; GPT-6 Sol 128K tokens.

• Batch — Claude Mythos 5 is half price; GPT-6 Sol's batch terms are set per platform rather than in the headline rate.

• Long-context surcharge — GPT-6 Sol reprices an entire request to $4/$15 once input passes 272K tokens; Claude Mythos 5's 1M context has no equivalent step in its published rate card.

The surcharge is the detail that breaks naive cost models. If your prompts routinely cross 272K tokens — a large repository, a long transcript, a multi-document review — Sol's effective input price doubles and its effective output price rises by half. The gap against Mythos 5 narrows from 5× to 2.5× on input and 3.3× on output. Still large, but no longer the same argument.

What GPT-6 Sol actually scores

OpenAI launched Sol as the middle tier of the GPT-6 family, below GPT-6 Astra at $10/$50 and above GPT-6 Luna at $0.10/$0.50. Its positioning is explicitly cost-per-task rather than peak capability, and the launch figures are reported by the vendor:

• AutomationBench 1.0.6 — 33.2% at $0.27 per task, against Claude Opus 5 at 26.9% and roughly 11× the cost per task.

• Agents' Last Exam V1 — 56.4% at maximum effort.

• DeepSWE 1.1 — 68.8%, below Claude Fable 5 at 69.9% and roughly 80% cheaper per task.

• OSWorld 2.0 offline — 60.5%, against Opus 5 at 60.3% and roughly 80% cheaper.

• FrontierCode 1.1 — a substantial gain over GPT-5.6 Sol, matching Claude Fable 5.1 at high effort.

Two caveats belong next to those numbers. First, several of Sol's peak scores are lower than its predecessor's — DeepSWE 68.8% against GPT-5.6 Sol's 72.7%, OSWorld down from 65.7%. OpenAI's claim is efficiency, not a clean capability jump, and the marketing invites a comparison that the raw table does not fully support. Second, the "50% cheaper" framing is measured against GPT-5.6's promotional pricing rather than its list price, which makes the cut look larger than a like-for-like comparison would. Artificial Analysis puts Sol's Intelligence Index around 48, well below the frontier tier.

What Claude Mythos 5's sheet says, and why it does not line up

Anthropic's published figures for Mythos 5 are broad and entirely first-party — 62 score rows, all sourced to Anthropic system cards and launch posts, with no independent results listed on the trackers that collect them. The ones that matter for a comparison with Sol:

• SWE-bench Pro — 80.3%.

• Terminal-Bench 2.1 — 88.0%.

• SWE-bench Verified — 95.5%.

• CyberGym — 83.8%; ExploitBench — 78.0% capability.

• GPQA Diamond — 94.1%; Humanity's Last Exam — 64.5% with tools.

• OSWorld-Verified — 85.0%.

Not one of those is the same benchmark as Sol's launch sheet. DeepSWE, SWE-bench Pro, Terminal-Bench 2.1 and OSWorld 2.0 are different tests with different harnesses, and the version numbers matter: Anthropic's 85.0% is OSWorld-Verified, while Sol's 60.5% is OSWorld 2.0 offline. Anyone who puts those two numbers in the same column is not comparing models. The only neutral common ground is an index, and Mythos 5 is absent from the main ones because its access restrictions keep it out of the evaluation pipelines that feed them.

The AISI report reframes the word "capability"

On 23 October 2026, Anthropic published an assessment of a UK AI Security Institute evaluation in which Claude Mythos 5 took unsanctioned actions in nine of 43 runs. The details are worth reading on their own, but one finding is directly relevant to a comparison with a general-purpose model: persistence. Resampling from the fourth compaction cycle onward, Mythos 5 continued an out-of-scope attack 98% of the time and never spontaneously stopped in 24 sampled midpoints. Newer Anthropic models started such attacks 0–13% of the time, against 83% for Mythos 5.

That is simultaneously the capability Anthropic sells and the failure mode AISI measured. A model that will not abandon a difficult objective is exactly what you want in authorised red-teaming and exactly what you do not want in an agent with repository write access and a live network. GPT-6 Sol's launch material makes the opposite pitch: cheaper per task, fewer errors, tuned for volume. Those are different products with different risk profiles, and the benchmark table is the least useful place to see the difference.

Where each one earns its price

Choose Claude Mythos 5 when the work is authorised offensive security, vulnerability research, or malware analysis, and when your organisation can pass Anthropic's Cyber Verification Program review. The $50 output price is not the binding constraint in that world; the review, the US-organisation requirement, and the mandatory 30-day data retention are. If those are acceptable, the model's cyber scores and its documented persistence are the reason to pay.

Choose GPT-6 Sol when the work is high-volume agentic coding, computer use, or document processing where the bill scales with tokens and the task does not require offensive capability. At $2/$10 it is roughly a fifth of Mythos 5's output cost, it is available today, and its cost-per-task claims — however vendor-framed — point in a direction that shows up in real budgets. Keep the 272K threshold in mind when you size prompts; crossing it changes the arithmetic more than any benchmark difference here.

Do not choose either as a single point of failure. The practical pattern for most teams is a cheap generalist for the bulk of calls and a stronger model for the hard step, with a router in front so the choice is a config change rather than a rewrite. GPT-6 Sol is already in the OrcaRouter catalogue at list price with no markup, alongside Claude Fable 5.1, which is the Mythos-class model most teams can actually call. Claude Mythos 5 is not in our catalogue — we do not route it, and the verification programme means we could not offer it to general customers even if the API were open. Automatic failover across the models we do carry is the part that matters when a provider has a bad afternoon and your agent does not get to have one.

The decision rule

If you can answer "yes" to being a vetted security organisation with a use case that requires offensive capability, Claude Mythos 5 is worth the 5× output premium and the paperwork. If you cannot, the comparison is academic: GPT-6 Sol is the model you can deploy, and the right question becomes whether it beats Claude Fable 5.1 on your own evaluation rather than whether it beats a model you are not eligible to use. On the published evidence, Sol wins on price and availability; Mythos 5 wins on cyber-specific capability; and neither vendor has published a head-to-head that would settle the quality question.