Qwen3.8 Max briefly took the #1 spot on Artificial Analysis's Agentic Index this week, the benchmark that scores how well a model handles multi-step tool use and long agent runs rather than one-shot chat answers. The headline going around right now — Qwen3.8 Max now ranked as the best overall model by agentic index — was accurate for about ten minutes on release day, before Anthropic's Claude Opus 5 reclaimed the top spot once Artificial Analysis finished re-running its grading pass. As of today, it sits tied for second at a score of 58, one point behind Opus 5's 59, and it's the strongest fully open-weight model anywhere near the top of that board. That distinction — open weights near the very top of an agentic leaderboard usually dominated by closed API-only models — is the actual story here, more than a rank that swapped within an hour of launch.
Short answer: Qwen3.8 Max hit #1 on Artificial Analysis's Agentic Index for a short window on August 3, 2026, then dropped to a near-tie for second (58, versus Claude Opus 5's 59) once rankings settled. It's the top open-weight agentic model available and costs $2/$6 per million tokens, but Kimi K3 scores close behind for less, and DeepSeek is far cheaper if budget matters more than the top score.
Disclosure: AISagely may earn a commission if you sign up through some links on this page, at no extra cost to you. That doesn't change what I recommend — I say plainly when a cheaper or free option is the smarter pick.
I ran the same multi-step coding brief through Qwen3.8 Max, Kimi K3, and DeepSeek's V4 Flash this week — refactor a small Python script, then use the terminal to run its tests and fix what broke — and cross-checked every price and score below against each vendor's own docs, not aggregator screenshots. This isn't a ranking of who wins a single benchmark; it's which model is actually worth paying for once you weigh the score against what a mistake or a slow run costs you.
How I picked these models
I only compared models that are either live on Artificial Analysis's Agentic Index today or directly relevant to the story — the models people are actually asking about after seeing this headline. For each one, I pulled current API pricing straight from the vendor's own docs, not a reseller or a benchmark tracker's summary, and I ran the same agentic coding task through the three I had direct API access to: Qwen3.8 Max, Kimi K3, and DeepSeek V4 Flash.
Score alone didn't decide the picks. A model that's one point ahead but costs 30% more per finished task, because it takes more steps to get there, isn't automatically the better buy. I weighed the Agentic Index score against real per-token pricing and whether the model ships open weights, since that changes who it's actually useful for.
Top picks at a glance
| Model | Best for | Price (official, per 1M tokens) | Free or open option |
|---|---|---|---|
| Qwen3.8 Max | Best overall — top open-weight agentic model | $2 in / $6 out (reported at GA, Aug 3, 2026) | Free via Qwen Chat; weights open-sourced |
| Kimi K3 | Best value — close score, lower cost per task | $3 in / $15 out, $0.30 cached | Free chat app at kimi.com |
| DeepSeek V4 Flash | Best free option — cheapest usable API | $0.14 in / $0.28 out, $0.0028 cached | Free chat app; open weights |
| Claude Opus 5 | Current #1 on the Agentic Index | $5 in / $25 out | No unlimited free tier; Pro from $17/mo |
Prices shift fast in this category. I confirmed DeepSeek's and Kimi's rates directly on DeepSeek’s API pricing docs and Moonshot’s Kimi K3 pricing page on August 7, 2026. Alibaba's own Model Studio pricing page hadn't added a Qwen3.8 Max row when I checked it the same day — the $2/$6 figure above is what multiple outlets reported at general availability, so treat it as recent but not yet vendor-page-confirmed.
Best overall: Qwen3.8 Max
Qwen3.8 Max is the pick if what you actually want is the best open-weight model for agent work, not just the model with the single highest number on one leaderboard. It's a 2.4-trillion-parameter mixture-of-experts model with a roughly 1-million-token context window, and Alibaba is releasing the weights, which no other model in the current top five on the Agentic Index is doing.
In my testing, Qwen3.8 Max finished the refactor-then-fix-tests task correctly, but it took noticeably more tool calls than Kimi K3 did to get there — which matches what Artificial Analysis found separately on its related Intelligence Index: it needs around 64 steps to Claude Opus 4.8's 14 for a comparable task, which is why its real per-task cost runs higher than its per-token price alone suggests. If you want the model closest to the top of the agentic board and you care about open weights, it's the right call. If you just want the highest score at the lowest cost, keep reading. See best AI models if you want the wider picture beyond just agentic benchmarks. Try Qwen here: Qwen.
Best for value: Kimi K3
Kimi K3 is the model I'd actually default to for agent workloads right now. On Artificial Analysis's Intelligence Index, it scores 57 to its 56, and it does it at roughly 25% less per finished task — $0.86 versus $1.14 — because it needs fewer steps to land the same result, per an August 6, 2026 comparison. Moonshot's own pricing confirms $3 per million input tokens and $15 per million output, dropping to $0.30 on cached input, which is competitive with it once you account for the step count.
When I tested it on the same coding brief, Kimi K3 finished in fewer turns and didn't need a follow-up prompt to catch a broken import the other two models missed on the first pass. It doesn't have open weights yet — Moonshot has said public weights are coming but hasn't shipped them — so it's not the pick if open-source matters to you specifically. For raw score-per-dollar on agentic tasks today, it's the strongest of the three. Try Kimi here: Kimi.
Best free option: DeepSeek V4 Flash
DeepSeek V4 Flash is the one to reach for if price is the deciding factor and you don't need the absolute top score. Official pricing is $0.14 per million input tokens and $0.28 per million output — a fraction of a cent for most real coding tasks — with cache-hit input at $0.0028, confirmed on DeepSeek’s own API docs. The chat app at chat.deepseek.com is free with no card required, and DeepSeek publishes open weights for its models.
It's not competing at the top of the Agentic Index the way Qwen3.8 and Kimi K3 are, and in my test it needed a second prompt to finish the test-fixing half of the task cleanly. For high-volume, lower-stakes agent work — batch scripts, routine refactors, anything where an occasional re-prompt is cheap — the price gap more than makes up for the score gap. My Qwen vs DeepSeek piece and how to use DeepSeek guide cover setup and a closer head-to-head if this is the direction you're leaning. Try DeepSeek here: Deepseek.
How to choose the right one
Start with what actually breaks your workflow: a wrong answer, or a slow one. If a bad agent run costs you real time to catch and fix, pay for the model closest to the top of the Agentic Index — that's currently Claude Opus 5 at $5/$25 per million tokens, with Qwen3.8 close behind and open-weight if you need to self-host or fine-tune.
If you're running high volume, lower-stakes agent tasks — the kind where re-running a failed step costs pennies, not your afternoon — Kimi K3's score-per-dollar or DeepSeek's near-free pricing make more sense than paying for the top score you won't consistently need. And if you're not sure yet which category your work falls into, that's a real answer too: start on a free tier (Qwen Chat, Kimi's chat app, or DeepSeek's chat app all cost nothing) and only move to the API once you know what a mistake actually costs you. See DeepSeek vs ChatGPT for how a budget open-weight model compares against a mainstream paid one on ordinary tasks, not just agent benchmarks.
One caution: leaderboard rankings like this one move fast. Artificial Analysis re-ran its Agentic Index grading the same week it launched and the rank changed within the hour. Check the live board before you buy on the strength of a headline — including this one.
My verdict
Qwen3.8 Max earns the "best overall" claim in one specific sense: it's the strongest fully open-weight model near the top of an agentic leaderboard that's normally dominated by closed models, and that's genuinely new. It is not, as of today, the single #1 model on the Agentic Index — that's Claude Opus 5, by one point.
For most people picking a model for real agent work, I'd start with Kimi K3 for the best score-per-dollar, keep DeepSeek V4 Flash on hand for high-volume cheap tasks, and reach for Qwen3.8 specifically when open weights matter to you — self-hosting, fine-tuning, or just not wanting to depend on one vendor's API. Don't buy on the strength of a "now ranked #1" headline alone; check how we test AI tools and the live board before you commit budget.
Frequently Asked Questions
Is Qwen3.8 Max really ranked as the best overall model by agentic index?
It was, briefly — for about ten minutes after its August 3, 2026 release, before Artificial Analysis re-ran its grading and Claude Opus 5 retook the top spot. As of today, it's tied for second at 58, one point behind Opus 5's 59, on the Agentic Index.
What makes Qwen3.8 Max different from Claude Opus 5 or Kimi K3?
It's the only model in the current top few on the Agentic Index shipping open weights, alongside a roughly 1-million-token context window. Claude Opus 5 currently scores higher; Kimi K3 scores close behind it at a lower cost per finished task.
Are the free versions of these models good enough?
For casual use and testing, yes. Qwen Chat, Kimi's chat app, and DeepSeek's chat app are all free with no card required. For production agent workloads where consistency matters, the paid APIs are worth it once you're running tasks daily.
How much should I expect to pay?
Anywhere from near-$0 to $25 per million output tokens depending on the model. DeepSeek V4 Flash is the cheapest at $0.14/$0.28 per million tokens; Qwen3.8 Max runs $2/$6; Kimi K3 is $3/$15; Claude Opus 5, the current Agentic Index leader, is $5/$25.
Does Qwen3.8 Max have a free tier?
Yes — Qwen Chat is free to use in a browser or the app with no card required, and Alibaba has open-sourced the model's weights, so it's also free to self-host if you have the hardware.