Qwen3.8-Flash-Next vs Claude Opus 4.8: What 33x the Price Actually Buys

Claude Opus 4.8 lists at $5.00 per million input tokens and $25.00 per million output on A​nthropic's own rate card. Qwen3.8-Flash-Next is A​libaba's open-weight preview of the Qwen4 architecture — weights on Hugging Face since August 24, 2026, on ModelScope since August 26 — and in the served Qwen3.8-Flash form it costs $0.15 and $0.47 on those same units. That is 33x on input and 53x on output for two models that, on the only public index that has scored both, sit within about two points of each other. The matchup looks lopsided and is not. What the multiplier actually buys is narrower and more specific than the gap suggests, and finding out which part of it your workload needs is the whole decision.

One thing to clear before anything else, because it shapes every number below: nobody has run these two head to head at matched settings on a named harness and published the result. What exists is each model's own score on the Artificial Analysis Intelligence Index, and those entries were evaluated on different dates, so treat them as adjacent rather than as a controlled comparison. Everything else here comes from the two vendors' published material and from independent harnesses that have actually run one of the two.

Claude Opus 4.8, from A​nthropic's own documentation

Claude Opus 4.8 was released on May 28, 2026, and A​nthropic now files it as a legacy model: still available, still sold, but no longer the head of the Opus line. A​nthropic's own model overview says you should consider migrating to Claude Opus 5.5, and the migration guide for that move is linked from the 4.8 page itself. As a fact about the model rather than a mark against it: legacy here means superseded at the top, not withdrawn.

What it carries: a 1M-token context window and 128K max output, raised to 300K on the Message Batches API under a beta header. Text and images in, text out, with vision, tool use, JSON and reasoning among its supported capabilities. Adaptive thinking with the effort parameter defaulting to high — a different default from Opus 5.5, which ships at medium, so an identical `effort` value does not mean the same thinking budget across the two. A reliable knowledge cutoff and training cutoff that both land in January 2026, five months behind the current Opus generation. Pricing of $5.00 per million input tokens, $25.00 per million output, $0.50 per million cache reads, $6.25 per million five-minute cache writes, $10.00 per million one-hour cache writes, and a flat 50% Batch API discount on input and output. A​nthropic commits to keeping the model available until at least May 28, 2027.

It is not a museum piece, either. The model ID `claude-opus-4-8` is live on the Claude API, A​mazon Bedrock, G​oogle Cloud, M​icrosoft Foundry and Claude Platform on AWS, and A​nthropic's own release material for Opus 5.5 documents that most cybersecurity requests in that newer deployment are routed down to Opus 4.8 by design. A model that is the designated fallback for the current flagship is still being asked to do real work.

Qwen3.8-Flash-Next, from A​libaba's own card

Qwen3.8-Flash-Next is not a Qwen4 model, and the distinction is A​libaba's own rather than a reading of it. The model card describes an experimental preview of the architecture that will underpin Qwen4 — the architecture shipped early, deliberately, so runtimes and quantisations are ready when the family arrives. It stores roughly 180B parameters in BF16: a 125B mixture-of-experts body activating 6B per token, a separate 51B n-gram embedding table, and about 4B of multi-token-prediction layers, across 48 layers with 512 experts of which 10 are routed and one is shared. The attention stack is the genuinely new part — 36 linear-attention layers running Gated Delta Network against 12 full-attention layers running Qwen Sparse Attention, the same roughly three-to-one hybrid split the wider Qwen3.8 generation uses.

Native context is 262,144 tokens, extensible to 1,000,000. Input is text, image and video; output is text. The licence is Qwen Community License 1.0, not Apache 2.0 — a real difference from the family's 27B, and the kind of thing that decides a procurement question months after it decides a benchmark argument. A​libaba positions the open weights as the experimental preview and Qwen3.8-Flash, served through Qwen Cloud, as the official production version with a default 1M context and built-in tools. Whether those are the same weights under different serving configuration is not stated on the card.

The numbers, labelled by who produced them

A​libaba's model card is a vendor table and is labelled as one here. On the card's own language evaluations, Qwen3.8-Flash-Next reports 91.9 on LiveCodeBench v6, 91.7 on GPQA Diamond, 81.0 on SWE-bench Multilingual, 62.5 on SWE-bench Pro, 58.7 on DeepSWE 1.1, 73.9 on CoWorkBench and 55.7 on JobBench. On the card's own comparison rows it loses to a Claude model on HLE, 35.9 against 40.0, and on NL2Repo-Bench, 48.1 against 54.2. Two caveats come from the card itself: several results use in-house benchmarks, and the HLE row notes it was judged by GPT-4o, which is worth knowing before quoting that particular figure to anyone.

A​nthropic publishes no equivalent public table for Claude Opus 4.8 using those same harnesses, so a vendor-to-vendor row-by-row comparison does not exist and should not be implied from either side's sheet. What does exist is a third party that has run both. Artificial Analysis carries Claude Opus 4.8 at 41.8 on its Intelligence Index — 23rd of the models it ranks, 84th percentile — and Qwen3.8-Flash-Next at 40, fifth of the 113 models in its comparison class. Both figures are the same provider's index, which makes them the closest thing to a common yardstick available today, with the caveat stated above about evaluation dates.

The rest of the independent picture, all of it from Artificial Analysis and all of it for the open-weight preview rather than the served Flash variant:

• Knowledge work — Qwen3.8-Flash-Next posts AA-Briefcase 1587, GDPval-AA v2 1647 and 56% on AutomationBench-AA, the agentic workflow benchmark built with Zapier on a private held-out task set.

• Terminal and scientific work — 25% on Terminal-Bench v4.0 and 51% on SciCode, against a much weaker showing on long-horizon repository tasks.

• Reasoning — Humanity's Last Exam 38% for the Qwen preview against 48.7% for Claude Opus 4.8 on the same provider's harness. On GPQA Diamond the two are effectively level: 91.7 on A​libaba's card against 92.0 for Opus 4.8 on Artificial Analysis.

• Factual recall versus confabulation — AA-Omniscience at minus 10 for the Qwen preview, the measure that penalises confident invention. No equivalent figure is published for Opus 4.8 on that harness, so the two cannot be compared on it.

• Speed — Artificial Analysis measures the Qwen preview at 53 output tokens per second with a 2.80-second time to first token, and calls it very verbose: 108k output tokens per task, of which 72k are reasoning tokens. Claude Opus 4.8, measured through our own playground over the last seven days, returns 66.7 output tokens per second at a 4.2-second median first token. Different harnesses, so treat the rates as indicative rather than as a race result.

• Context — 262,144 native and 1,000,000 extended for Qwen3.8-Flash-Next against 1,000,000 native for Claude Opus 4.8.

Read as a group rather than one line at a time, the pattern is consistent with the price gap being misleading in both directions. The cheap model is genuinely close on knowledge work, terminal tasks and general reasoning — closer than a 33x price difference would predict. The expensive model is ahead where the work is long, ambiguous and expensive to get wrong, and it has a 1M-token context natively rather than by extension.

The price question, on the unit that decides it

Per token, the gap is stark and it is arithmetic on published numbers. On the served Qwen3.8-Flash, input is 33x cheaper and output 53x cheaper than Claude Opus 4.8. Caching narrows it slightly in A​nthropic's favour relative to its own input rate — Claude Opus 4.8 cache reads run at $0.50 per million, one-tenth of its input price — but the served Qwen3.8-Flash carries a cache-read rate of $0.0184 per million on our own published list, which is under 4% of the Opus 4.8 equivalent. For a long-lived prompt with heavy re-reading, that difference compounds in the same direction as everything else.

Per completed task is the number that survives token verbosity, and here the honest answer is that the comparison is incomplete on one side. Artificial Analysis publishes a cost-per-task figure for the models it runs; the preview's own verbosity — 245M tokens burned on the index run against a 140M median — means its per-token advantage shrinks at the task level rather than disappearing, and no matched-setting run of these two exists to settle it. If you need one number to plan with, model it yourself on your own traffic instead of borrowing someone else's token counts.

Where each one is the right call

Claude Opus 4.8 is the right choice when the work is high-stakes and single-pass: long-horizon agentic runs where a wrong branch is expensive to unwind, complex multi-stage debugging across a large codebase, and anything already built against the Claude thinking and effort controls, where an identical `effort` setting and a January 2026 knowledge cutoff are known quantities rather than unknowns. It is also the safer pick when you want a vendor commitment with a date on it and a model that is still the documented fallback for its successor's hardest traffic.

Qwen3.8-Flash-Next is the right choice when the workload is wide rather than deep — high-volume classification, extraction, screenshot and document understanding, agentic sweeps where you would otherwise batch for cost — and when you want to evaluate the Qwen4 architecture on your own data months before Qwen4 exists. Its strong independent signals are on knowledge work and terminal tasks; its weak signals are on long-horizon repository generation and one-shot agent pass rates, which is exactly where production code changes live. Run a low-stakes copy of a real workload against it before it touches anything that matters.

The rule, stated as a rule: pick the model whose failure mode you can afford. If a wrong answer costs you a retry, the 33x is free money. If a wrong answer costs you a rollback at 3am, the cheaper model has not yet earned that decision — and the fact that it is an experimental preview with a community licence rather than a released product line is part of the reason why, independent of the scores.

Running both without a second contract

Both ends of this comparison are live on one catalogue. Claude Opus 4.8 is routed as `anthropic/claude-opus-4.8` at $5.00 and $25.00 per million tokens — A​nthropic's published list price passed through with 0% markup, so nothing here changes if that rate card does. The served production variant of the Qwen side is routed as `qwen/qwen3.8-flash` at $0.15 and $0.47, also at the provider's list price, on a 1M-token context with text, image and video input.

The open-weight preview itself is a different matter and worth stating plainly: Qwen3.8-Flash-Next is not one of our routes. To use the preview you download the weights from Hugging Face or ModelScope and serve them yourself — vLLM supports it from day one with an offload path that keeps the 51B-parameter n-gram embedding table in host memory rather than permanent VRAM, SGLang and TensorRT-LLM are in N​VIDIA's Day 0 list, and Ollama's library carries an MLX build for Apple Silicon. What we route is the production variant, which is the one you can actually call today.

The practical arrangement for a team that wants both is a traffic split rather than a decision: one key, one SDK, one bill, with a percentage of requests sent to the cheap model and automatic failover absorbing the moments when an experimental architecture or an aging one misbehaves. That is a smaller engineering commitment than the framing of this matchup suggests, and it is how most of the useful answers to "which one" actually get found — by running both against real traffic for a week and reading the results.

What would change this page

Three things, in order of how likely they are to actually happen. A matched-setting evaluation of Qwen3.8-Flash-Next against Claude Opus 4.8 on a named harness would replace every hedged sentence above with a number. The arrival of the full Qwen4 family would move the preview from "the architecture, early" to "the previous generation of the architecture", which changes what it is for. And any pricing move on the Opus line would reopen the multiplier at the centre of this page — the served Qwen3.8-Flash rate passed through on our side moves the same day the provider moves it, so the comparison stays current even when the numbers do not.

Until then, the single fact that decides this pairing most often is not a benchmark: it is that the cheap side is an experimental preview you serve yourself and the expensive side is a legacy model you rent with a support relationship and a 2027 availability commitment. Teams that need to own the stack and can absorb an unproven architecture should take the 33x and test it properly. Teams that need the answer to be right the first time should pay for the model whose failure mode they have already characterised.