Gemini 4 Carbon vs Fable 5.2: Qwen4-Max preview has no score and retires November 25

The two-model question already has a page. Gemini 4 Carbon vs Claude Fable 5.2 for agents prices a 50,000-in, 15,000-out turn at $0.25 on Carbon through November 30, 2026, and $1.00 on Fable 5.2. On November 11, Alibaba opened a third id, qwen4-max-preview, and described it as the Qwen family’s next flagship for coding and agentic work. The docs give it a 1 million token context window, tool calling, and a retirement date. They do not give it a benchmark score or a dollar price per million tokens.

That empty column is this page. Carbon is $2 and $10 per million tokens through November 30. Fable is $8 and $40. Qwen4-Max preview is a Token Plan credit against a standard rate Alibaba has not printed, and the id is scheduled to disappear at 24:00 on November 25, 2026 (UTC+8).

What did Alibaba open on November 11?

Alibaba Cloud Model Studio, last updated November 11, 2026, and the matching QwenCloud page, read November 12, describe the same model. It generates text, reads images and video, and reasons. The intended work is coding, office productivity, frontend design, and complex agentic tasks. Tool calling covers function calling, web search, and built-in tools. Structured output covers JSON object and JSON schema. The APIs listed are OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages.

The spec block on both pages:

Source: aiagentsinfo.com, from the Alibaba Cloud and QwenCloud Qwen4-Max-Preview pages read November 12, 2026.

Two base URLs, both official, for two consoles. Alibaba Cloud’s English page says Token Plan Personal and https://token-plan.ap-southeast-1.maas.aliyuncs.com. QwenCloud says Token Plan Individual and https://token-plan.maas.qwencloudapi.com. A key from one console will not explain itself on the other host. QwenCloud’s Token Plan Individual page adds a billing alias: qwen4.0-max-preview. Both ids work. Bills show qwen4.0-max-preview.

24:00 on November 25 (UTC+8) is 16:00 UTC on November 25. A coding agent that still has this id in its config on the morning of November 26 in London is past the retirement line the docs print.

The output-budget note is the harness footgun. Starting with this model, max_tokens means the same thing as max_completion_tokens. Both cap thinking and the answer together. A limit sized for the answer alone can cut the reply off, because thinking spends the same budget. Older Qwen models keep their previous behavior. The page also says not to set reasoning_effort and thinking_budget together. On Chat Completions, preserve_thinking is on by default. Historical thinking goes back in reasoning_content. The page says not to fold it into content.

The directory listing keeps the id, the window, and the retirement date in one place.

How do Carbon, Fable 5.2, and Qwen4-Max preview compare?

Every number below was on a vendor page this desk opened on November 12, 2026. A blank cell is a blank cell. Alibaba’s Qwen4 pages and the model-prices page do not print a FrontierSWE score, a FrontierCode score, an AutomationBench score, or a dollar rate for qwen4-max-preview.

Source: aiagentsinfo.com. Qwen columns from the Alibaba Cloud and QwenCloud pages read November 12, 2026. Carbon’s prices from Google’s November 10 availability post, including the November 30 footnote. Carbon’s token pool from the Enterprise model page read November 12. Fable’s price, 1M context, and 128K max output from the model overview read November 12. The three benchmark rows from Anthropic’s November 9 announcement table, read November 12. Fast mode’s 2.5 times claim is Anthropic’s, from that announcement.

Read the FrontierCode row as two different effort settings. Fable’s 49.1% is marked Xhigh. Carbon’s 41.1% is marked Medium. Those cells do not answer “who wins at the same effort.” The decision between those two ids, the HTTP 400s, and the rows this table leaves out are on Gemini 4 Carbon vs Claude Fable 5.2 for agents. The wider set, including GPT-6 Astra, is Claude Fable 5.2 vs GPT-6 Astra. Qwen4-Max preview is on neither scoreboard.

What does one turn cost, where a price exists?

A single turn of 50,000 input tokens and 15,000 output tokens, with no cache and no extra thinking tokens piled on top of the 15,000:

Source: aiagentsinfo.com. Carbon and Fable rates from the pages named above. The products are this desk’s arithmetic. Cache, batch, and fast mode are outside the three dollar figures. Qwen’s cell is blank because the rate is blank.

Carbon is one quarter of Fable’s list price on this turn through November 30, and half of it after that. The AI agent cost calculator applies published per-token rates. It cannot price an id whose vendor has not published one. Use the calculator for Fable. Use the table for Carbon until the rate card in the calculator matches the November 10 post. Leave Qwen off a cost-per-task chart.

The blank is sharper once thinking is on. Qwen’s default is xhigh, and thinking shares the 393,216 output cap. A harness that logs “15,000 output tokens” on this model has not logged 15,000 answer tokens. Carbon’s Enterprise page puts thinking in the output pool as well. A fair cost-per-task eval counts those tokens. This page will not invent a thinking-token average for any of the three.

How do you point a coding agent at it?

QwenCloud’s Claude Code page, read November 12, is the concrete wiring. The Token Plan block is:

export ANTHROPIC_BASE_URL="https://token-plan.maas.qwencloudapi.com/apps/anthropic"
export ANTHROPIC_MODEL="auto"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="qwen3.6-flash"
export ANTHROPIC_DEFAULT_SONNET_MODEL="qwen3.8-flash"
export ANTHROPIC_DEFAULT_OPUS_MODEL="qwen4-max-preview"
export CLAUDE_CODE_SUBAGENT_MODEL="auto"
export CLAUDE_CODE_MAX_CONTEXT_TOKENS="983616"

The same page says: when using qwen4-max-preview, set CLAUDE_CODE_MAX_OUTPUT_TOKENS=64000 if long output is truncated. The model’s own cap is 393,216. The env var they tell you to set is 64,000. The context env is 983,616, against a model window the spec pages write as 1M. Those are harness caps on top of the model caps. A loop that assumes the full 393,216 will be cut at the env var.

ANTHROPIC_MODEL is auto. The preview id is in the Opus slot. A session that never asks for Opus is on whatever auto selects, which this page does not claim is qwen4-max-preview.

The pay-as-you-go block on the same Claude Code page maps Opus to qwen3.7-max. It does not mention qwen4-max-preview. That matches the model page: the preview is Token Plan only. Claude Code pointed at a pay-as-you-go Qwen key is a different model.

QwenCloud’s Token Plan nav also lists Codex, Qwen Code, Qoder, OpenClaw, and others. This page read the Claude Code instructions and did not re-copy every client. Use the page for the client you run. The Individual plan’s feature line says the subscription works with mainstream programming and agent tools. The cheapest plan’s concurrency cap is still 1–2 agents, in the table below.

Why a credit is not a token price

Token Plan Individual, read November 12, bills every model and built-in tool from one credit balance. The Qwen4-Max preview benefit on that page, and on the team page, is the same sentence: credits are as low as 20% of the standard rate, which QwenCloud calls 5x usage, and calls from 22:00 to 08:00 (UTC+8) take an additional 60% off, which the page says is 8% of the standard rate. 0.20 × 0.40 = 0.08, so the night figure matches the sentence. The denominator, credits per token at the standard rate, is not on the model-prices page. That page’s only mentions of Qwen4-Max-Preview are the previous and next links.

Source: aiagentsinfo.com, from QwenCloud’s Token Plan Individual page read November 12, 2026. The page does not print an end date for the new-user prices in the block this desk copied.

The November 9, 2026 price-reduction notice, effective 20:00:00 (UTC+8) that day, is why Pro is $43. QwenCloud’s table shows the previous website price at $68 and 180,000 credits, and the new price at $43 and 200,000 credits. The notice calls the Pro cut 37%. (68 − 43) / 68 is 36.8% to one decimal place. Existing Pro subscribers keep the old price until the next cycle. The credit top-up applies to the current cycle, and credits already used stay used. Lite, Essential, and Standard kept their dollar prices and gained credits: 11,500 to 20,000, 25,500 to 45,000, and 45,000 to 80,000.

Source: aiagentsinfo.com, from QwenCloud’s Token Plan team page read November 12, 2026. The page says unused credits in a team credit pack expire at the end of that pack’s month.

When the Individual quota and any credit packs are gone, the overview says the service pauses until the next subscription month. A harness that fans out subagents hits the concurrency column before it hits the model. Lite is 1–2 concurrent agents. A tree of eight workers belongs on Pro, at 6–8, or it queues on a smaller plan. That cap is a product limit, separate from the model’s tool-calling support.

None of those dollar figures converts into a cost per task for qwen4-max-preview until QwenCloud prints credits per token. A Lite plan at $6 for 20,000 credits is $0.30 per thousand credits. It is not a statement about how many tokens one credit buys on this model.

What should a coding-agent harness do before November 25?

Keep the Carbon versus Fable choice on the agent comparison: Carbon when the job looks like AutomationBench or the bill is the constraint, Fable when the job is coding or computer use and your eval needs the gap on Anthropic’s table. Re-check Carbon on December 1. The coding-agent comparison is the product layer. Claude and Gemini are the families.

Use qwen4-max-preview as a timed eval beside that choice, not as the production default. Point Claude Code’s Opus slot at the Token Plan Anthropic base URL, keep max_tokens large enough for xhigh thinking plus the answer, and set the 64,000 output env if the client truncates. Run your own tasks before 16:00 UTC on November 25. Record the credit burn, because Alibaba has not given you a rate to predict it. Remove the id when the preview ends. The pages say the model is retired, and they say this preview does not change while it is up, so a score you measure this week is a score on a frozen preview, not on a model you will still be able to call in December.

An empty benchmark cell is an empty cell. It does not place Qwen above Carbon or Fable, and it does not place Qwen below them. The harnesses topic is where the loop around any of these three models belongs.