Flagship models, freshest numbers. Intelligence = Artificial Analysis Index / headline benchmark. Price = USD per 1M tokens (in / out).
| Model | Flagship intel | Price in / out | Access | Sharpest at |
|---|---|---|---|---|
Claude Opus 4.8 Anthropic |
61.4 Intelligence Index — #1 |
$5 / $30 | Closed | Coding (SWE-bench Pro 69.2%), computer-use agents |
GPT-5.6 Sol OpenAI |
94.6% GPQA Diamond — #1 |
$5 / $30 | Closed | Reasoning, enterprise work, "ultra" subagent mode |
DeepSeek V4 Pro DeepSeek |
~59 Closes gap on reasoning |
$0.15 / $3.48 | Open | 1.6T MoE (49B active), 1M context, price-per-IQ |
Kimi K2.7-Code Moonshot AI |
58.6 SWE-Bench Pro (K2.6) |
~$0.95 / — | Open | Agentic coding, Agent Swarm (300 sub-agents), token efficiency |
↳ No model wins every column. Closed labs lead raw capability; open Chinese models lead on cost-per-capability by a wide margin.
Three tiers from workhorse to budget. New max reasoning effort and an "ultra" mode that spins up subagents for complex work. Also shipped GPT-Live full-duplex voice models that speak & listen at once.
Native 1M-token context, promo pricing through Aug 31. Opus 4.8 remains the intelligence-index leader (61.4). Cowork expanded to mobile + web; Claude for Government in beta.
Two open MoE models: V4-Pro (1.6T / 49B active) and V4-Flash (284B / 13B). Both 1M context. Claims it outstrips GPT-5.2 & Gemini 3.0 Pro on some reasoning tasks — at a fraction of the price.
Cuts reasoning tokens ~30% vs K2.6, +21.8% on its own code bench. Open-weights (Modified MIT). K2.6's Agent Swarm scales to 300 sub-agents / 4,000 steps.
The frontier has split into two games. Capability is a coin-flip between Anthropic and OpenAI — Opus 4.8 leads the composite index, GPT-5.6 Sol leads the hardest reasoning benchmark. Value is owned outright by the open Chinese labs: DeepSeek and Kimi deliver 80–90% of frontier quality at up to 70× lower cost, and now hold four of the top-five open-weight slots. This week's real story isn't a new smartest model — it's that agents hit the cost wall, so every lab is racing on efficiency and subagent orchestration.
For high-volume, non-critical calls (drafts, classification, internal tools), route to DeepSeek V4 / Kimi — 80–90% quality at a fraction of the cost. Reserve Opus/GPT-5.6 for client-facing quality.
"Ultra"/Agent-Swarm modes are the new frontier. Design Code@ pipelines around many small delegated agents, not one big prompt — that's where speed & cost wins now live.
Default in Claude Code with 1M context and promotional pricing through Aug 31 — cheapest window to run large-repo agentic coding at frontier quality. Use it now.
Token-efficiency (Sol +54%, Kimi −30%) now moves margins more than benchmark points. Bench your Code@ workloads on cost-per-task, not leaderboard rank.