The AI Model Spec Sheet
Ten models, the fields that decide real choices — context, price, what each is actually good at, and the gotcha the pricing page doesn't mention.
Every model that matters, one page, real numbers. Ten models, the fields that decide real choices — context, price, what it’s actually good at, and the gotcha the pricing page doesn’t mention. Three honesty rules. One: these are list prices per million tokens; prompt caching and batch tiers routinely cut real bills 50–90%, so treat list price as the ceiling. Two: every row carries a verified date, because this table has a half-life measured in weeks — we re-check quarterly and whenever something big ships. Three: a † means we couldn’t verify that cell to a primary source on the verified date; we’d rather show you the hole than fill it with a guess.
Verified: July 6, 2026.
The table
| Model | Maker | Context | $ in /M | $ out /M | Best for |
|---|---|---|---|---|---|
| Claude Fable 5 | Anthropic | 1M | $10.00 | $50.00 | The hardest work — frontier reasoning, long autonomous coding runs |
| Claude Opus 4.8 | Anthropic | †(verify) | $5.00 | $25.00 | Heavy daily coding and agent work at half frontier price |
| Claude Sonnet 5 | Anthropic | 1M | $2.00* | $10.00* | The default — most production workloads |
| GPT-5.5 | OpenAI | ~1.05M (922K in / 128K out) | $5.00 | $30.00 | OpenAI-ecosystem frontier; long-document professional work |
| Gemini 3.1 Pro | †(verify) | $2.00 | $12.00 | Frontier reasoning per dollar; analysis-heavy work | |
| Gemini 3.5 Flash | †(verify) | $1.50 | $9.00 | High-volume professional writing at scale | |
| Grok 4.3 | xAI | 1M | $1.25 | $2.50 | Live X/social data + cheapest frontier-class API |
| GLM-5.2 | Z.ai | 1M | $1.40 | $4.40 | The self-host pick — strongest open weights (MIT) |
| DeepSeek V4-Flash | DeepSeek | 1M | $0.14 | $0.28 | Bulk work at pennies — classification, extraction, first-pass drafts |
| Qwen 3.7 Max | Alibaba | †(verify) | $1.25 | †(verify) | Budget mid-tier reasoning and coding |
* Introductory pricing through Aug 31, 2026; rises to $3.00 / $15.00 Sept 1. Budget at the September number; enjoy the summer one.
The rows, with the parts the table can’t hold
Claude Fable 5 — Anthropic
Strengths: The most capable model you can currently use — top of the intelligence composites and the best available score on the hardest coding benchmark (80.3% SWE-Bench Pro). Purpose-built for long-horizon autonomous runs, which is exactly the shape of work everything else still fumbles. Gotchas: Twice Opus 4.8’s price, so the job has to earn it. And a production consideration nobody’s pricing page mentions: it spent June 12–July 1 pulled offline under a US export-control order before returning with a tighter safety layer. Availability history is a spec — if your workflow can’t survive a vendor outage, design the fallback before you need it. Verdict: The ceiling. When the work is frontier-hard and failure is expensive, this is the hire — for everything else, look one and two rows down.
Claude Opus 4.8 — Anthropic
Strengths: The heavy-duty workhorse — holds Anthropic’s top SWE-bench Verified score and remains the favorite inside Claude Code and Cursor, at half the flagship’s price. Gotchas: Sonnet 5 now beats it on some agentic benchmarks (Terminal-Bench 2.1: 80.4 vs 74.6) at 40% of the price — so “Opus by default” is no longer automatic. Check your task’s shape before paying the 2.5×. Verdict: Still the daily driver for serious coding and agent work — but audit it against Sonnet 5 this quarter; the answer may have changed under you.
Claude Sonnet 5 — Anthropic
Strengths: Launched June 30 and immediately became the default on Claude’s free and Pro plans: native 1M context, finishes multi-step tasks earlier Sonnets stalled on, checks its own output unprompted, and posts frontier-adjacent numbers (63.2% SWE-Bench Pro) at mid-tier prices. Gotchas: The $2/$10 is introductory — it becomes $3/$15 on September 1, so any cost model built on the summer price has an expiry date in it. Verdict: The new default for most production workloads. If you were routing everything to Opus, this is the quarter to re-run that decision.
GPT-5.5 — OpenAI
Strengths: OpenAI’s frontier for complex professional work — ~1.05M context, strong reasoning scores (93.5% GPQA), deep ecosystem, and prompt caching that drops repeated input to $0.50/M, a 90% cut that matters enormously for agents. Gotchas: Two billing cliffs. Output is $30/M — six times input — so generation-heavy work is where the bill lives. And any prompt over 272K input tokens re-prices the entire session at 2× input / 1.5× output; long-document workloads should be split deliberately. Also note: the GPT-5.6 family (Sol/Terra/Luna) is in limited preview with Terra expected GA mid-July — this row may age in weeks. Verdict: The pick if you live in OpenAI’s ecosystem or need its tooling — with your context architecture designed around the 272K line, not discovering it on an invoice.
Gemini 3.1 Pro — Google
Strengths: The reasoning-per-dollar leader — 94.3% GPQA Diamond and the top ARC-AGI-2 score (77.1%, genuine novel reasoning, not memorization) at $2/$12, a fraction of what those numbers cost anywhere else. Strong multilingual. Gotchas: Google’s model-version churn is real; pin exact versions in anything production. Context window unverified at press time†. Verdict: The best analysis brain you can rent cheaply. For research, evaluation, and reasoning-heavy pipelines, this is the value call of the table.
Gemini 3.5 Flash — Google
Strengths: The price-performance frontier — top professional-writing marks (1,656 GDPval-AA Elo) at $1.50/$9, which is why it’s the pick for email, reports, and any writing done at volume. Gotchas: It’s the efficient tier, not the deep one — work needing frontier reasoning depth belongs a row up. Context unverified†. Verdict: The volume writer. If the workflow produces a lot of competent prose on a budget, stop shopping.
Grok 4.3 — xAI
Strengths: Shipped April 30 with the thing nobody else has — native real-time X data — plus reasoning on by default and one of the cheapest frontier-class APIs going ($1.25/$2.50, 1M context). Grok 4.20 stays in service as the 2M long-context option. Gotchas: The most permissive guardrails of any frontier model cuts both ways — run a brand-safety review before anything client-facing. Consumer access is staged through SuperGrok tiers; Grok 4.5 is already in private beta, so this row moves soon. Verdict: One job, done uniquely well: anything that needs the live social pulse. The bargain API pricing is the bonus.
GLM-5.2 — Z.ai
Strengths: The strongest open-weight model, full stop — MIT license, 1M context, 62.1% SWE-Bench Pro, top open-weights intelligence scores. Hosted at $1.40/$4.40 or run it on your own metal. Gotchas: The ecosystem and tooling around it are younger than the big three’s; budget integration time. Hosted-API route reintroduces the data questions self-hosting exists to answer. Verdict: If the requirement is open weights — for data control, cost control, or principle — this is the one. Everything else open-weight is now competing for second.
DeepSeek V4-Flash — DeepSeek
Strengths: A 1M-context model at $0.14/$0.28 — the cheapest real capability per token on this page by an order of magnitude. This is what you point at classification, extraction, routing, and bulk first-pass drafts. Gotchas: Hosted-API data residency deserves a real look before client or sensitive data touches it; self-hosting is the control move. And cheap models make it painless to automate work that shouldn’t exist — the 100-times rule still applies. Verdict: The bulk-work engine. High-volume, low-stakes, verifiable-output jobs — nothing else on this page comes close on cost.
Qwen 3.7 Max — Alibaba
Strengths: The cheapest model in the global top ten by reasoning benchmark, and a legitimate mid-tier coding value. Gotchas: The 3.7 tier is closed-weight — self-hosting means stepping down to the 3.6 family — and the same data-residency homework as above applies for client work. Output pricing unverified at press time†. Verdict: The budget brain for cost-sensitive reasoning and coding — with the residency question answered before deployment, not after.
Near misses and in preview
Didn’t make the ten, worth knowing: GPT-5.4 ($2.50/$15) — OpenAI’s production workhorse at exactly half of 5.5; o4-mini ($0.55/$2.20) — the reasoning-value sleeper; GPT-4.1 nano ($0.10/$0.40) — the routing-and-extraction cheapskate that refuses to die.
In preview (not usable at scale yet, watch the dates): GPT-5.6 family — Sol $5/$30, Terra $2.50/$15, Luna $1/$6; Terra expected GA mid-July. Grok 4.5 — private beta since June 28. Claude Mythos 5 — restricted to approved organizations. When these go GA, this sheet gets its event re-verify.
How to read this page
The verified date at the top is the whole promise: every number was checked against provider lists and live trackers that day, and the page gets a full pass quarterly plus an event pass when anything major ships. List prices are ceilings — caching, batch, and flex tiers cut real costs dramatically, and the model that’s cheapest on paper isn’t always cheapest on your workload. And the standing rule of the whole site applies here hardest: there is no best model — there’s a best model per job. That’s not a hedge; it’s the finding.
Related
- Which agent runs these models — job by job → Claude Code vs Codex vs Cursor vs Copilot
- Costs compound at volume — run yours → Is It Worth Automating?
- Route the work to the right AI first → The Four Questions
Frequently asked questions
Are these the real prices I'll pay?
Treat them as the ceiling. These are list prices per million tokens; prompt caching and batch tiers routinely cut real bills 50–90%, so the model that's cheapest on paper isn't always cheapest on your workload.
What does the dagger mark mean?
A dagger means we couldn't verify that cell to a primary source on the verified date. We'd rather show you the hole than fill it with a guess — check the provider's price page before you rely on it.
Which model is best?
There is no best model — there's a best model per job. That's not a hedge; it's the finding. Match the row to the work: frontier-hard tasks up top, bulk high-volume work at the bottom, most production workloads in the middle.
How current is this table?
It carries a verified date because it has a half-life measured in weeks. Every number was checked against provider lists and live trackers on that date; the page gets a full pass quarterly plus an event pass whenever something major ships.