
Qwen 4 vs Qwen 3.8 Max: What Actually Changes on Your Bill When You Upgrade
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 90 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0717 tok/s
- OrcaNEWOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 87 tok/s
- OrcaNEWOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 133 tok/s
- OrcaNEWOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$4.00 / $20.00 per 1M tokens · 72 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 126 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 65 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 92 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 44 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 193 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 147 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
- OrcaOrca: OrcaCyber Zero 1.52026-10-10$3.00 / $7.50 per 1M tokens · 123 tok/s
- openaiOpenAI: GPT-6.1 Sol2026-09-2955Intelligence
- anthropicAnthropic: Claude Sonnet 5.52026-09-2855Intelligence
- typesafeTypeSafe: Jev 1.132026-09-24$0.04 / $0.00 per 1M tokens · 219 tok/s
The migration from Qwen3.8-Max to Qwen4-Max is unusual among flagship upgrades because the price moves in your favour and the capability moves at the same time. Qwen3.8-Max has been generally available since 3 August 2026 at $2.00 in and $6.00 out per million tokens with a one-million-token context. Qwen4-Max, which replaced it at the top of the line on 16 November 2026, is $1.50 and $4.50 for the same window, scores 43 against 34 on the Artificial Analysis Intelligence Index, and costs $3.27 per index task against $10.02. A 25 per cent cut on both sides of the meter, nine index points, and a third of the cost per completed task — the kind of upgrade that normally arrives with a price rise attached. The catch, and there is always one in a same-vendor migration, is that the output budget got smaller even as everything else got bigger.
The four numbers that moved, and the one that moved backwards
Upgrades within a single vendor are easier to evaluate than cross-vendor comparisons because the tokenizer lineage, the API surface and the tool-calling conventions all stay put. What changes is a short list, and you can check it against your own usage in an afternoon.
• Price — $2.00/$6.00 down to $1.50/$4.50 per million tokens, a 25 per cent cut on both sides
• Cached read — $0.25 down to $0.075 per million, a 70 per cent cut
• Independent score — 34 up to 43 on the same index, same board, both read 17 November 2026
• Cost per index task — $10.02 down to $3.27, a 67 per cent reduction
• Shared thinking-plus-answer budget — 262,144 tokens, down from a larger allowance on the 3.8 generation
• Active parameters — 71 billion, down from 95 billion, on the same 2.4-trillion-parameter backbone
• Context — unchanged at 1,000,000 tokens
• Weights — neither flagship publishes any, though Qwen3.8-Max's open-weight sibling Qwen3.8-27B does
Two of those deserve more than a line. The active-parameter reduction is the architectural claim behind the rest: Qwen4-Max activates 71 billion parameters per token where Qwen3.8-Max activated 95 billion, and Alibaba attributes the efficiency gain to a new architecture rather than to pruning — gated DeltaNet layers interleaved with a sparse attention mechanism that indexes context at micro-block granularity, four parallel residual branches with dynamic read and write gates, and a context-based N-gram embedding table of 328 billion parameters that expands capacity without expanding per-token compute.
The shared budget is the one that will surprise people. Qwen4-Max pools its reasoning and its answer into a single 262,144-token allowance. If you were generating 150,000-token outputs on Qwen3.8-Max with thinking enabled, you now have 112,144 tokens left for thinking — and on a hard problem, thinking will eat that. The failure looks like a truncated answer, not an error, which is the worst way for a limit to surface.
Why the per-task cost fell by two thirds when the token price fell by a quarter
The token prices explain a quarter of the saving. The other two thirds come from the model doing less work to reach the same answer, and this is the part of the launch that is hardest to verify and most important to the business case.
Alibaba's own numbers: on its Work360 suite, token usage falls about 24 per cent at xhigh reasoning effort and 35 per cent at high; on Terminal-Bench 4.0, 48 and 49 per cent. Average cost per task on Work360 drops from $4.43 to $1.67, a 62 per cent reduction. Those are vendor figures, measured by the vendor, against the vendor's previous model, with no independent reproduction — and they are directional evidence that the efficiency claim is not just a price cut wearing a costume.
The independent corroboration is narrower but agrees. Artificial Analysis measures $10.02 per index task for Qwen3.8-Max and $3.27 for Qwen4-Max. Nobody published $4.43 or $1.67 except Alibaba; the $10.02 and $3.27 are somebody else's arithmetic on somebody else's workload, and they point the same way.
Put a number on your own work. A pipeline doing 5,000 tasks a day at 30,000 input and 4,000 output tokens, no caching, on the published US-dollar rates: Qwen3.8-Max bills $50 of input and $120 of output per day, sum $170. Qwen4-Max bills $37.50 and $90 — $127.50. About $1,275 saved a month before caching, and considerably more if a stable prefix lets you exploit the $0.075 cache read rather than the $0.25 one.
The efficiency story has a second half nobody quotes
Cheaper tokens per task is only half of a business case. The other half is elapsed time, and here the picture is murkier than the price table.
Artificial Analysis measures Qwen3.8-Max at 59 output tokens per second and 48 minutes per index task — slow, and if you are running long-horizon agentic work that number is the one that governs your nightly window. Qwen4-Max publishes no output speed and no per-task timing at all. It uses roughly half the tokens on some benchmarks, which would imply roughly half the wall clock at equal throughput, but nobody has published the throughput.
So the honest framing of the upgrade is: you will definitely pay less per token, almost certainly pay less per task, and probably finish faster. Only the first of those three is a fact you can put in a contract today. The second is supported by an independent measurement. The third is inference.
That matters for scheduling decisions specifically. If your 48-minute index task at 59 tokens per second is already at the edge of what fits overnight, you need a pilot with your own prompts before you migrate — not because Qwen4-Max is likely to be slower, but because a shared 262,144-token thinking budget changes the shape of a long run in ways a benchmark average will not reveal.
What did not change, and why that is the best news here
The context window is the same million tokens. The modality surface is the same — text, image and video in, text out. The API surface is the same, with thinking controlled by enable_thinking or by reasoning.effort on the Responses API, and the model served across Alibaba's regions in OpenAI-compatible, Anthropic-compatible and native formats.
For anyone already on Qwen3.8-Max this is what makes the upgrade cheap: the migration is a model identifier change, not a rewrite. No prompt restructuring, no tool-calling schema to renegotiate, no context-management change. The two things you must re-test are the ones the numbers flag — the output budget under thinking, and whatever your latency-sensitive paths are.
It also means the sensible pattern is shadow traffic rather than a switch: send a slice of production requests to both, compare outputs and token counts, and let the shared 262,144-token budget reveal itself on the long cases before it does so in front of a customer.

Should you keep the 0902 snapshot?
Qwen3.8-Max has a dated snapshot, qwen3.8-max-0902, pinned to 2 September 2026 with the same capabilities and pricing as the base model. Snapshots exist for reproducibility, and the case for keeping one is strongest exactly when you are about to change the underlying model.
If you have an evaluation suite that you trust, run it against the snapshot before you retire the alias. The snapshot is your control group; the alias was never stable by design, and Qwen4-Max is a different model, not a version bump of the same one. Two details worth checking in that run: whether your long-output cases survive the smaller shared budget, and whether your cost-per-task actually moved by the margin the price table predicts. If it did not, the token efficiency claim is not holding on your workload, and the case for migrating narrows to the price cut alone — still worth having, but a quarter rather than two thirds.
One more option that the flagship comparison tends to hide: Qwen3.8-27B is open-weight under Apache-2.0 and self-hostable, and we serve it as qwen/qwen3.8-27b. For high-volume, low-complexity work the arithmetic is completely different from either flagship, and "upgrade from 3.8-Max to Qwen4-Max" is not the only move available. If your workload is classification, extraction or short-form generation, a 27B dense model you can run yourself may beat both on unit cost — and it is the one option in this family you can pin to a version and audit.
The migration, in order
First, inventory your outputs. Any path that requests more than about 120,000 output tokens with thinking enabled is at risk under the shared 262,144 budget. That is a grep of your call sites, not a benchmark.
Second, measure your own token consumption per task on both models for a representative week. The published efficiency figures are between 24 and 49 per cent depending on benchmark and effort, which is a wide band — wide enough that your own number could sit outside it.
Third, time your p95. Qwen4-Max has no published throughput, so your first latency number will be the first latency number anyone outside Alibaba has.
Fourth, decide on the open-weight tier. If part of the volume can move to qwen/qwen3.8-27b, the flagship migration only has to carry the hard cases, and the blended cost falls further than either single-model plan suggests.
Running both flagships behind one endpoint is the low-effort version of this: qwen/qwen3.8-max stays at $2.00/$6.00 and a Qwen4-Max request is a routing rule rather than a new contract, so you can shift traffic gradually and fail back the moment the shared budget bites. Worth saying plainly that Qwen4-Max itself is not among the 215 identifiers we currently serve — the upgraded model routes through Alibaba's own API — so what you get from us today is the incumbents, priced at list with nothing added.
The verdict, which is unusual
Migrate, unless you generate very long outputs. A same-vendor upgrade that cuts the price by a quarter, cuts the cache price by 70 per cent, and adds nine index points is close to a free lunch, and the per-task evidence — from the vendor and from an independent board — both say the real saving is larger than the token price implies.
The exception is narrow and specific: workloads whose output regularly exceeds roughly 120,000 tokens with thinking on. For those, the shared 262,144-token budget is a capability regression dressed as an efficiency gain, and the correct move is to stay on the 3.8 generation or split the request.
What to watch next is the number that is missing rather than the ones that are published. A model that is two-thirds cheaper per task and has no measured throughput is a strong bet with an unquantified risk, and the first independent timing of Qwen4-Max will settle more of this argument than any further benchmark score.

The part of this family you can run yourself
One advantage of a same-vendor upgrade that does not exist in either flagship: the tier below them is open and portable. Qwen3.8-27B ships under Apache-2.0 with text, image and video input and a 262,144-token window, and we serve it as qwen/qwen3.8-27b — on our own infrastructure, at list price with no markup, pin-able in a way neither flagship is. For a migration whose main risk is a shared output budget you cannot audit, having a versioned open-weight fallback in the same family is the cheapest insurance available.

Compared in this article1
Detected from this article · Benchmarks: Artificial Analysis · updated daily
