25 popular paid models · updated 2026-08-23 · cache-heavy agentic profile

LLM Cost & Capability

Data last refreshed 2026-08-28 · refreshes automatically every week

The 25 most-used paid cloud models, repriced on a cache-heavy agentic workload (the pattern typical of coding agents and long-running tools, where ~96% of tokens are cached re-reads). Two columns carry the decision: Cached-in — the rate for repeated tokens — and Eff $/M, the blended cost per million on this profile. Models with a recent price change show the old rate struck out, the new rate in green if it dropped or red if it rose.

• 1.740.435 price dropped
• ↑ Sep 1 price rising
• Eff $/M in teal = blended cost
• ↕ click any column header to sort

• Updated 2026-08-23 against primary publishers. Recent changes: Sol on short-context promo $5/$30→$4/$20 (through Nov 21); Sonnet 5 intro $2/$10 rising +50% to $3/$15 on Sep 1; DeepSeek V4-Flash cached-in confirmed at $0.0028.
• Eff $/M = blended cost per 1M tokens on a cache-heavy agentic profile (0.1% input · 0.3% output · 3.6% cache-write · 96% cache-read). Multiply by your own monthly token volume (in millions) to estimate spend.
• * cached-in rate estimated at ~10% of input (provider doesn't publish one); the rest are official. The cached-in rate dominates cost on a cache-heavy workload — DeepSeek's is the lowest.
• Prices per 1M tokens, standard tier, USD. Earlier drops shown: DeepSeek V4-Pro (−75%, permanent) and GPT-5.6 Luna (−80%, Jul 30).

Best model by use case

Each use case is ranked by the benchmark that governs it — score on the left, effective cost on the right. The row in teal is the best value: the cheapest model that still clears the quality bar for that job. Example weights reflect a dev-heavy workload.

Use case	Governing score	Example weight
Production coding	SWE-bench Verified	75%
Content generation	LMArena (preference)	13%
Research & analysis	Reasoning index	7%
Bulk / unattended	Cost-first	fleet
Sensitive / private	Local-only	on-device

• ‡ Luna's coding score is contested — one leaderboard shows 93.0%, another 73.3%. SWE-bench Verified is saturated (top models within ~4 pts), so numbers swing by source. † Kimi K3's 93.4% is self-reported.
• Content & research rankings are directional (tier-based from LMArena standings), not exact per-model scores.

As of August 2026 · "Popular" = major paid models across leading providers · figures are API list-price value per 1M tokens. Sources: OpenAI, Anthropic, Google, xAI, DeepSeek, Alibaba, Moonshot, Z.ai, Mistral, Amazon, Cohere, Meta published pricing; SWE-bench & LMArena leaderboards. Prices change often — verify before relying on them.