Model comparisons
Which model, and what it costs
to finish the job.
Every pair I have run, indexed by model. Published price, the independent score, and the number that decides it in practice: what one finished task cost, not what a million tokens list for.
30 comparisons across 22 models. Each one is rewritten when the numbers move, not left to rot.
Claude Opus 5.5 beats GPT-6 Astra on every score and still costs more to finish the job
Anthropic cut every price on Claude Opus 5.5. The bill against Opus 5 still went up.
Claude Haiku 4.5 is the cheap tier. GPT-6 Luna just made it the expensive one.
Pick the model you run
Models you can buy today. Each one lists what I have measured it against.
Claude Fable 5.1
Anthropic · 7 comparisonsClaude Sonnet 5
Anthropic · 7 comparisonsClaude Haiku 4.5
Anthropic · 3 comparisonsClaude Opus 5.5
Anthropic · 3 comparisonsClaude Code
Anthropic · 1 comparisonClaude Mythos 5.1
Anthropic · 1 comparisonGPT-6 Astra
OpenAI · 6 comparisonsGPT-6 Sol
OpenAI · 2 comparisonsCodex
OpenAI · 1 comparisonGPT-6 Luna
OpenAI · 1 comparisonGemini 3.8 Flash
Google · 2 comparisonsMuse Spark 1.3
Meta · 1 comparisonGrok 4.6
xAI · 1 comparisonDeepSeek V4.1 Flash
DeepSeek · 1 comparisonModels that have been replaced
Still here, because people run what they already bought. Each article says what took its place.
GPT-5.6 Luna
Claude Sonnet 4.6
GPT-5.6 Terra
Claude Mythos 5
Claude Opus 4.8
In the order I ran them
Newest first, with the verdict on one line.
PairWhat it comes down toDate
GPT-6 AstravsClaude Opus 5.5better and dearer per task28 Sept 20268 min readClaude Opus 5.5vsOpus 5cheaper per token, dearer per task23 Sept 20269 min readGPT-6 LunavsClaude Haiku 4.5a tenth of the price, twice the score23 Sept 20267 min readGPT-6 SolvsClaude Opus 5.5the index and the invoice disagree23 Sept 20268 min readGPT-6 SolvsGPT-6 Astraties on code, loses the terminal23 Sept 20267 min readClaude CodevsCodexcoding agent benchmark and cost16 Sept 20267 min readClaude Fable 5.1vsFable 5what changed, what breaks16 Sept 20267 min readGPT-5.6 LunavsClaude Fable 5.115 points, 50x price16 Sept 20266 min readGPT-5.6 TerravsClaude Fable 5.111 points at 5x price16 Sept 20266 min readClaude Haiku 4.5vsGPT-5.6 Lunafive times dearer, 20 points behind16 Sept 20266 min readThe systems behind these numbers are the actual work.
Model choice is one decision inside a revenue system. The rest of what I build and run is in the field notes.
Read the field notes →