Qwen’s new 0902 snapshot is available through QwenCloud. The useful story is not an unqualified benchmark victory: it is a claimed coding and Cowork uplift paired with unusually cheap repeated-prefix reads, wrapped in a release that still needs independent testing.
Three takeaways
- The change is a hosted revision. QwenCloud lists
qwen3.8-max-0902, aliasqwen3.8-max-2026-09-02, with a 1M-token context, 991K maximum input, 131K maximum output, and 15K RPM on the displayed plan. QwenCloud’s model page is the operational source. - Qwen reports broad gains, not independent proof. Its comparison graphic shows 0902 ahead of the previous Qwen3.8-Max on every displayed row with both values present, but the new column is not independently listed on the benchmark owners’ current leaderboards.
- The cache price is the deployment hook. Standard input is $2 per million tokens, while an implicit-cache hit is $0.25 and an explicit-cache read is $0.17. Those discounts apply to eligible cached input; reasoning, output, retries, tools, tax, and foreign-exchange costs remain.
What changed in Qwen3.8-Max-0902?
Qwen announced the revision on 2 September 2026, describing further post-training for Coding and Cowork and stronger performance on complex enterprise work, scientific research, and long-horizon workflows. The official Qwen announcement presents 0902 as an upgrade to the Qwen3.8-Max model rather than a new parameter-scale generation.
QwenCloud now lists the exact model ID and the alias qwen3.8-max-2026-09-02. The catalog exposes text, image, and video input, thinking mode, function calling, structured output, web search, code execution, and context caching. Its OpenAI-compatible documentation also lists 0902 as a supported model ID, although teams should still pin the exact ID and region-specific endpoint they intend to use. QwenCloud API documentation
The 2.4T headline is consistent with Alibaba’s August announcement for the base Qwen3.8-Max, which described a sparse Mixture-of-Experts design. That earlier release also described 95B activated parameters. The 0902 announcement confirms 2.4T but does not separately restate 95B, so EyesTech is not presenting the active-parameter figure as independently verified for this revision. Alibaba’s base-model announcement
The benchmark uplift is real as a claim—and bounded as evidence
Qwen’s supplied comparison graphic reports the following selected results:
| Evaluation | 0902 | Previous Qwen3.8-Max | Claude Opus 5 in the graphic | What the row measures |
|---|---|---|---|---|
| TerminalBench 3.0 | 29.0 | 11.3 | 42.7 | Terminal-agent task result |
| DeepSWE 1.1 | 69.3 | 56.6 | 73.6 | Long-horizon software engineering |
| NL2Repo-Bench | 64.9 | 55.9 | 72.3 | Repository generation from requirements |
| QwenSWEBench V2 | 70.0 | 55.1 | 68.0 | Qwen’s complex real-world software benchmark |
| CoWorkBench | 76.1 | 74.8 | 79.6 | Long-horizon office work |
| WorkArena Elo | 1468 | 1348 | 1437 | Qwen’s expert-graded collaboration score |
The table is useful because the direction of change is consistent: the 0902 column improves on the previous version across coding, office-work, and collaboration rows. But it is not a controlled league table. The graphic mixes resolution rates, benchmark percentages, in-house evaluations, and Elo. Several rows use an agent harness, long timeouts, and large context windows; the footnotes also say competitor numbers can be the best published result across harnesses.
The underlying benchmark projects are legitimate, but the release-column evidence is still vendor-reported. Terminal-Bench is versioned; DeepSWE and JobBench expose different public configurations; and QwenSWEBench V2 and CoWorkBench are Qwen’s own evaluations. The WorkArena number should not be confused with the canonical ServiceNow WorkArena browser benchmark. Terminal-Bench · DeepSWE · JobBench · ServiceNow WorkArena
The editorial conclusion is therefore narrower and more useful: Qwen reports a broad 0902 uplift, but readers need a matched harness, fixed task set, latency, token count, retries, and human-cleanup data before treating it as a production winner.
The price story is cache geometry, not just $2 per million
QwenCloud lists this rate card:
| Billed item | Price per 1M tokens | How to read it |
|---|---|---|
| Input | $2.00 | Uncached input |
| Output | $6.00 | Includes billed reasoning/output tokens under the selected mode |
| Implicit cache hit | $0.25 | Cached input when the service detects a reusable prefix |
| Explicit cache creation | $2.50 | First write of an explicit cache |
| Explicit cache read | $0.17 | Reuse of an eligible explicit cache |
That makes an implicit hit 87.5% cheaper than ordinary input and an explicit read 91.5% cheaper. The savings are only on the cached input portion. Output is still three times the ordinary input rate, and QwenCloud documents thinking as enabled by default for the Qwen3.8 series. QwenCloud’s cache documentation explains the billing boundary and the short explicit-cache window.
Here is a practical calculation for an agent that repeatedly carries the same codebase or policy manual:
- Stable prefix: 850K tokens
- Changing task input: 50K tokens
- Output: 100K tokens
| Mode | Calculation | Cost per call |
|---|---|---|
| No cache | 850K × $2 + 50K × $2 + 100K × $6 | $2.40 |
| Implicit hit | 850K × $0.25 + 50K × $2 + 100K × $6 | $0.9125 |
| Explicit creation | 850K × $2.50 + 50K × $2 + 100K × $6 | $2.825 |
| Explicit read | 850K × $0.17 + 50K × $2 + 100K × $6 | $0.8445 |
If ten calls fit inside the explicit cache’s documented five-minute validity window, one creation plus nine reads costs about $10.43, versus $24.00 without caching—a calculated saving of roughly 56.6%. That is a scenario, not a guarantee. Cache hits can miss; long-running agents can outlive the window; and retries, tool calls, reasoning expansion, and human review can dominate the bill.
What this changes for India-based teams
The API is worth a controlled trial for teams with large stable prefixes and repeated calls: repository agents, document-analysis pipelines, internal policy assistants, and research workflows. Start with a fixed fixture and log:
- exact model ID and endpoint;
- input, output, and reasoning tokens;
- cache-hit status and cached-token count;
- time to first token, total latency, retries, and tool calls; and
- task success plus human cleanup.
Do not read the 1M context label as durable memory. The page lists 991K maximum input in standard mode and 131K maximum output, while an explicit cache is a short-lived billing optimization. A long-horizon workflow still needs its own state, retrieval, checkpointing, and failure recovery.
For India specifically, Alibaba’s public Model Studio region list names China Beijing, Singapore, Germany, Japan, Hong Kong, and the United States; India is not listed. That does not prove Indian access is unavailable, but the reviewed documentation does not establish India-hosted inference or India data residency. Prices are in USD, so the landed cost is the list price multiplied by the actual settlement FX rate, plus applicable tax and payment fees. Alibaba Cloud regions
Who should care
- Builders: trial 0902 now if your workflow can exploit repeated context and you can capture task-level telemetry.
- Engineering leaders: treat $0.17–$0.25 as a cache-input rate, not a total cost-per-action forecast.
- Procurement and governance teams: clarify endpoint region, data handling, account limits, SLA, tax, and model-version pinning before moving sensitive workloads.
- Researchers: use the graphic to identify hypotheses, then reproduce the tasks under a fixed harness.
Limitations and update note
No authenticated EyesTech API run was performed. No public 0902 p50/p95 latency, throughput, independent cost-per-success result, or separate 0902 open-weight checkpoint was verified. The benchmark values above are transcribed from Qwen’s comparison graphic and labelled as such; they should not be presented as independent leaderboard results.
Recheck the QwenCloud model page, cache rules, region availability, and price before committing to a production budget. The next useful update for the Model Reality Index is a matched run of 0902, Qwen3.8-Max, and two peers on the same coding and long-context fixtures.
Try the model: Qwen3.8-Max-0902 on QwenCloud
Suggested internal context: EyesTech’s analysis of China’s flash-model economics and the cache-first comparison of Claude Fable 5.1.
