Qwen3.8-Max vs GPT-5.6 Sol: Terminal-Bench and Price

You are paying frontier rates to run a terminal coding agent, and if that API line is a big enough number to argue about, August 3 gave you something to argue with: Alibaba published a Terminal-Bench 2.1 score two points under GPT-5.6 Sol's at $2 per million tokens in and $6 out. Here is the shortest route to an answer you can defend to whoever signs off on the spend.

Four steps get you to a decision before any independent benchmark lands

None of this requires waiting for Artificial Analysis or LMArena. All four steps can be done in an afternoon, and the first two are arithmetic.

• Step 1. Put the two price sheets next to each other, not a comparison table. Qwen published its numbers directly in the launch post: $2.00 per million input tokens, $6.00 per million output, and $0.25 per million on implicit cache hits. Sol's rates appear in neither Qwen's launch materials nor Grok's summary of them, so read them on OpenAI's own API pricing page rather than trusting a third-party grid, this one included.
• Step 2. Price your actual traffic, weighted toward output and cache. A coding agent spends most of its tokens on repeated context and generated patches, because the loop reads a repo, plans, edits, runs tests and then re-reads the diff. Take last month's real input and output token counts out of your provider dashboard and multiply. At Qwen's published rates, 20M input plus 4M output is $40 plus $24. If the same volume at Sol rates is not several times that, the rest of this post is optional.
• Step 3. Get the exact model ID and check the endpoint shape before you plan any migration work. The launch post points to the model page on qwencloud.com and to Qwen Studio at chat.qwen.ai for a no-setup trial. Take the model string from that page rather than guessing at something like qwen3.8-max. Then find out whether the endpoint is OpenAI-compatible. That single fact is the difference between a one-line base-URL change in Aider or your own harness and a week of adapter work around a different tool-calling format.
• Step 4. Replay five real tasks from your own backlog through both models, on your own harness. Terminal-Bench 2.1 is the benchmark Qwen chose to lead its comparison table with, and 86.6 describes performance on that fixed task set under a scaffold Qwen configured. It says nothing about your build system, your test runner, or the tool definitions your agent actually carries. Pick five tickets you already closed, so you know what the correct diff looks like, and read the results against the 86.6 rather than instead of it.

The published numbers, all of them from Qwen's own comparison table as relayed in Grok's summary of the launch:

Qwen's own framing of what the model is for sets expectations well above a benchmark table:

Autonomous coding: 10+ days of self-evolving development, from empty folder to production without hand-holding, complete project trace in the GitHub.

That claim comes with a repository link, which is more than most long-horizon claims ship with. It is also the part of the announcement that no benchmark in the table measures.

The comparison breaks in five specific places

Every number in that table came from the vendor

Qwen ran these evaluations and Qwen chose which columns to print. Self-reported evaluation is standard across the industry, and it still carries the weight of a vendor claim rather than a measurement. Terminal-Bench in particular is harness-sensitive: the scaffold, the retry policy and the timeout all move the result, and a lab evaluating its own model configures all three. When an outside evaluation lands, expect the 86.6 to move in one direction or the other. Commit a migration on 86.6, watch an independent run come back several points lower, and you are the one explaining the gap to someone who read the tweet.

The SWE-bench Pro line contradicts the Terminal-Bench line

Qwen3.8-Max loses Terminal-Bench to Sol by 2.2 points and wins SWE-bench Pro by 3.1. Both are coding benchmarks, and they disagree about which model is better. Read that as evidence the two models have different shapes rather than as a tie. The more useful comparison in that row is vertical: Fable 5 scores 80.0 on SWE-bench Pro against Qwen's 67.7 and Opus 4.8's 69.2. Beating GPT-5.6 Sol on one benchmark is a long way from reaching the top of the field, and a table that only prints the rows you win is a table worth reading twice.

The $0.25 cache rate only lands if your prefix stays stable

Implicit caching at $0.25 against a $2.00 base rate is an 87.5% discount on cached input, which is the single largest cost lever on the page. It only applies to a prefix the provider recognizes as repeated. Agent harnesses that reshuffle tool definitions, inject a changing timestamp near the top of the system prompt, or re-order retrieved file context between turns will miss the cache on most calls. What it looks like when it happens: your projected bill from Step 2 was built on cache hits, and the actual invoice comes in near the full $2.00 input rate. Check your provider's reported cache-hit ratio in week one, not month two.

Open weights at 2.4 trillion parameters are not weights you can run

The announcement promises Qwen3.8-Max weights next week alongside Qwen3.8-27B, and the two releases land in completely different places. Take the 2.4T parameter count from the launch post and quantize it to 4 bits and you get roughly 1.2TB before you count anything else, which is arithmetic off the parameter count rather than a hardware figure Qwen published. Either way it puts the Max in a data center, not on a workstation with a couple of consumer GPUs. If your reason for caring about open weights is self-hosting, the 27B is the release that matters to you and it will not score 86.6 on anything. If your reason is licensing durability or the ability to run inside a customer's VPC on rented accelerators, the Max weights are relevant and the cost is the cluster.

No license has been named

The launch post says "open weights" and stops there. It does not say Apache 2.0, it does not say MIT, and it does not say custom-commercial-with-a-revenue-clause. What it commits to is a date, next week, and two models going open. Terms for either are absent. Until the weights land with a LICENSE file, treat the open-weights promise as a plan rather than a fact you can build a deployment strategy on. The same applies to the independent early testing mentioned in Grok's summary, which puts Qwen3.8-Max around 80 against Kimi K3's 83 on architecture tasks: no methodology, no tester named, no harness disclosed. That is a rumor with a number attached.

Two conditions make this switch the wrong move regardless of the numbers

The first is a harness you did not write. If your agent workflow lives inside a product that decides which model runs, the model comparison is not yours to act on. Terminal-Bench scores are irrelevant to a tool that ships one backend, and the cost of leaving a coding environment your team already knows dwarfs a per-token difference you cannot access, which is where a lot of developers sit right now.

The second is data residency and vendor review. Qwen3.8-Max served through Alibaba's cloud means your prompts, which for a coding agent means your source code, transit an Alibaba endpoint. Some organizations resolved that question years ago and some cannot start the conversation. If yours is the second kind, the open-weights release next week is the version of this decision worth waiting for, because it moves the question from "do we approve this vendor" to "can we afford this cluster". I would wait the week either way before rebuilding anything, since the license terms and the first independent Terminal-Bench run should both land inside it.

If your monthly spend on coding-agent tokens is under about $100, skip every step above. The arithmetic in Step 2 is the whole argument, and at that volume a switch from Sol's rates to $2/$6 saves you less than one afternoon of re-tuning prompts, re-checking tool-call behavior and discovering which of your five replayed tasks now fails. Come back to this when the bill is large enough that a third of it is worth a day of your time.