Claude Fable 5Anthropic (US)
Proprietary- Positioned for this workload
- Cited tool-calling record
- $10/$50 per 1M tokens in/out
$10.00 in / $50.00 out per 1M tokens, verified 2026-07-15
Evidence
AI Inference Guide
Model Advisor
Pick a model the way deployments actually get decided: hardware and constraints first, then capability, with every claim linked to the leaderboard or spec page it came from. From a phone NPU to an 8x H200 node, the advisor names the model, the quantization, the serving stack, and how many GPUs the memory math really needs.
Start from a common setup
Where will it run?
What is the main job?
Context you actually need
Model origin
License
Data sovereignty
Also worth benchmarking
GPT-5.6 SolOpenAI (US)
Proprietary- Positioned for this workload
- Cited tool-calling record
- $5/$30 per 1M tokens in/out
$5.00 in / $30.00 out per 1M tokens, verified 2026-07-15
Kimi K3Moonshot AI (China)
Kimi K3 License- Positioned for this workload
- Cited tool-calling record
- MIT-like, but Model-as-a-Service businesses over $20M/year need a commercial agreement, and products over 100M MAU or $20M monthly revenue must display "Kimi K3" attribution.
Claude Sonnet 5Anthropic (US)
Proprietary- Positioned for this workload
- Cited tool-calling record
- $2/$10 per 1M tokens in/out
$2.00 in / $10.00 out per 1M tokens (Introductory rate through 2026-08-31; standard pricing is $3/$15.), verified 2026-07-15
Claude Haiku 4.5Anthropic (US)
Proprietary- Positioned for this workload
- $1/$5 per 1M tokens in/out
$1.00 in / $5.00 out per 1M tokens, verified 2026-07-15
Side-by-side comparison
Compare all candidates side by side, with the full sizing table. Pro feature.
Discuss this shortlist with the AI expert
The expert gets your requirements and this shortlist as context. You review the question before sending.
Dataset verified . Estimates, not promises: validate with a load test at your real prompt and output lengths. Leaderboards move weekly; every claim links to its source so you can re-check it. See the model spec table, what each benchmark proves and capacity planning for the underlying method.
How the advisor decides
Hardware first
Practitioners size the model to the memory budget before anything else, so the advisor starts from where it runs: hosted API, Apple unified memory, a workstation card, a datacenter fleet (legacy V100S-class included, with its constraint set), a phone, or the browser.
The job, not the leaderboard
A benchmark-topping chat model can still fail agent loops. Tool-calling maturity is scored separately from raw capability, and agentic work enforces a quantization floor (Q5 dense, Q6 MoE) because 4-bit degrades tool calls before it degrades prose.
Context and concurrency are memory
The KV cache grows with every token in flight: 50 users at 8K tokens can outweigh the weights themselves. The advisor prices that in per architecture (GQA, and MLA where the model publishes it) instead of assuming the marketing context window is free.
Origin as a policy, not a mood
Three positions: no constraint, no Chinese-hosted APIs (self-hosted open weights keep data local), or no Chinese-origin weights at all. Each is a real posture in 2026 procurement; the advisor treats them as filters, not judgments.
License before capability
Apache 2.0 and MIT ship without strings. Community licenses carry MAU caps and competitor clauses that surface late in procurement. Every recommendation shows its license class and links the text.
Sovereignty is jurisdiction
EU residency on a US cloud is not sovereignty. The EU filter prefers EU vendors and self-hosted open weights; air-gapped rules out hosted APIs entirely.
The leaderboards behind the evidence
No single leaderboard answers "which model should I use". These are the boards the advisor cites, what each one actually measures, and the caveat its publisher is quietest about. Scores go stale in weeks; that is why every evidence chip carries the date it was read and links to the live board instead of pretending a snapshot is the truth.
Artificial Analysis Intelligence Index
Composite of nine independently re-run evals (v4.1: GDPval-AA, tau3-Banking, Terminal-Bench 2.1, SciCode, HLE, GPQA Diamond, CritPt, AA-Omniscience, AA-LCR), plus price, speed and latency across ~170 models.
Caveat: The index composition changes between versions, so scores are not comparable across v3/v4/v4.1. Cite the version.
Arena (Text)
Crowdsourced pairwise human preference, Elo-style, ~6M votes.
Caveat: Measures preference, not correctness, and the 2025 "Leaderboard Illusion" paper documented sampling advantages for big labs.
Arena (Agent)
Human preference over agentic task executions, the newest Arena sub-board.
Caveat: Young board; vote volume is far below the Text arena.
SWE-bench Pro (public set)
Resolve rate on 731 public hard repo tasks (avg 107 changed lines across 4 files); the contamination-resistant successor to SWE-bench Verified.
Caveat: Scores are agent+model systems, never the model alone, and Scale sells eval services to the labs it ranks.
Terminal-Bench 2.1
Agent+model systems on 89 terminal/CLI tasks under the standardized Harbor harness.
Caveat: The harness and agent choice are a large confound; version churn breaks longitudinal comparison.
tau2-bench
Agentic reliability under policy: simulated customer-service domains with tool calls and a simulated user, scored pass^1.
Caveat: The simulated user is itself an LLM, so scores partly measure the simulator. Near saturation; tau3 is the live frontier.
tau3-Banking
The hardest current tau domain (RAG + policy + tools); top scores sit below 50%, so it still discriminates.
Caveat: New (March 2026) and volatile; pre-v1.0.1 results are not comparable.
Berkeley Function Calling Leaderboard V4
Tool-call correctness (schema-exact), with cost and latency columns; broad open-weight coverage.
Caveat: Single-call accuracy does not transfer to multi-turn reliability; models ace BFCL and stumble on tau-style loops.
LiveBench
Monthly-refreshed contamination-hardened tasks: reasoning, coding, agentic coding, math, data analysis.
Caveat: Covers fewer models than the aggregators; third-party mirrors of it go stale badly.
AA Long Context Reasoning
Cross-document reasoning over ~100k-token real document sets, run continuously across ~170 models.
Caveat: One hundred questions; treat small deltas as noise.
GPQA Diamond
Graduate-level science questions; PhDs with internet access score ~65%.
Caveat: Saturated: frontier models cluster at 94-95% on 198 questions, so it is a floor check, not a differentiator.
Humanity's Last Exam
Frontier academic reasoning, 2,500 questions with a private held-out set.
Caveat: Academic-recall flavored; hallucination under uncertainty is severe on it.
OpenRouter Rankings
Revealed preference: live share of real token spend, split by task.
Caveat: Reflects OpenRouter’s user base, not enterprise; free tiers and price cuts spike rankings without capability change.
The memory math, shown in full
Weights are parameters times bytes per parameter: roughly 2 bytes at FP16, 1 at FP8, and 0.55 to 0.82 across the 4-to-6-bit GGUF range once format overhead is counted. The KV cache adds 2 x layers x KV heads x head dimension x bytes for every token in flight, which is why concurrency and context, not parameter count, decide most capacity plans. MoE models occupy VRAM at their total size but decode at their active size. The advisor applies a runtime overhead factor and an allocator headroom fraction on top, and still tells you to load-test before renting anything.
The full method, including hosting-spectrum and serving-framework choices, is in capacity planning.