Comparisons
Side-by-side comparisons of the models in the pool, worked out from list price, multiplier, context window and capabilities.
- Claude Sonnet 5 vs Opus 5: which one to runBoth carry a 1M-token context. Opus 5 lists at 1.67× Sonnet 5’s official rate, but at 0.25× against 0.35× it bills 1.19× as much on both input and output. Here is when that buys you anything.
- Cheapest models by task: what each kind of work actually costsThe lowest multiplier is not the lowest price. Effective rates from $0.050 per million input tokens, sorted by the constraint each workload puts on the choice.
- Claude Opus 5 vs GPT 6 Astra Stable: 4× the price for 5% more contextAstra takes 1,050,000 tokens to Opus 5’s 1,000,000 and costs $2.00/$10.00 per million against $0.50/$2.50. One is a pool lane at 0.1×, the other a dedicated lane at 0.2× on twice the official rate.
- Claude Opus 5 vs Opus 4.8: what actually changedSame 1M context, same 128K output ceiling, same $5/$25 official rate, same 0.1× multiplier. The catalog separates them by release date and nothing else.
- Claude Sonnet 5 vs GPT 5.6 Sol: price, context, and the switching costBoth bill at 0.1×. Sol costs 1.67× on input and 2× on output, and takes 372,000 tokens of context against Sonnet 5’s 1,000,000. They also speak two different request shapes.
- Claude Sonnet 5 vs Sonnet 4.6: context, output ceiling, priceIdentical $3/$15 rates on the 0.1× lane, but 1M context against 200K and a 128K output ceiling against 64K. On the stable lanes the pricing inverts.
- Gemini 3.1 Pro vs Gemini Flash: the price gap depends on which FlashPro costs 4x gemini-3-flash but only 1.33x gemini-3.8-flash on input. Same 1,000,000-token window, same capabilities, same route — so the comparison is entirely about rate and about what your work can detect.
- Gemini 3.8 Flash vs 3.7 Flash: same rate, twice the multiplierIdentical official rates of $0.75/$3.75 per million and identical specifications, but 3.8 Flash is billed at 0.2x and 3.7 Flash at 0.1x. That makes the newer id exactly 2x the price here.
- GPT 5.5 vs GPT 5.4: same window, 2× the rateBoth carry a 272,000-token context, a 64,000-token output ceiling and the 0.1× rate. GPT 5.5 lists at exactly 2× GPT 5.4 on both input and output. What the newer generation does not change.
- GPT 5.6 Luna at 0.3×: when the higher multiplier still costs lessLuna is billed at 0.3× while its two siblings are 0.1×, and it is still the cheapest per-token id in the GPT 5.6 family. The multiplier is not the price — here is the arithmetic.
- GPT 5.6 Sol vs Terra: the same envelope at 2.5× the priceIdentical 372,000-token context, identical 128,000-token output ceiling, identical 0.1× rate. Sol lists at 2.5× Terra on both input and output — here is where that actually shows up.
- Grok 4.6 vs Grok 4.5: what the point release changedSame 0.2x rate, same 500,000-token context, same $2/$6 per million official ($0.40/$1.20 billed), same 65,536-token output ceiling. The only figures that move are the release date and the cached-read rate.
- Image generation models compared: three ids, two billing modelsgrok-imagine-image at $0.002 an image and grok-imagine-image-quality at $0.008 are 0.2× of $0.01 and $0.04. gpt-image-2 is 0.1× but billed on tokens, not per image.
- Largest context models: who reaches 1M, and what it costs to use itFifteen catalog entries hold 1,000,000 tokens or more. Their effective input rates span 80×, from $0.050 to $4.00 per million, and the output ceiling splits the group in half.
- Thinking models compared: who has it, and what it costs across providers25 of 31 catalog entries list thinking, across all four lineups. There is no separate rate for it — output is billed at 3× to 6× the input rate, and that is where the cost lands.
- Tool calling compared: 27 models, three request shapesEvery text model in the catalog supports tools. What separates them is the route — /v1/messages, /v1/responses or /v1/chat/completions — and what a multi-turn loop costs at each price point.