Reinventing.AI

AI Agent Trends

AI Agents Trends: Claude Fable 5.2 Arrives 20 Percent Cheaper and Qwen4-Max Lists at $1.50 as Agent Running Costs Fall

Anthropic cut its top model to $8 and $40 per million tokens on November 9. Alibaba listed Qwen4-Max at $1.50 and $4.50. What that changes for agent costs.

AI Agent Insights Team6 min
AI Agent Model Pricing Comparison

Two model vendors lowered the price of their most capable models within one week. Anthropic released Claude Fable 5.2 on November 9 and, according to Anthropic's Fable product page, priced it at $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Fable 5.1. Alibaba opened a preview of Qwen4-Max on November 11, and Alibaba Cloud's Model Studio price list now shows the model at $1.50 per million input tokens and $4.50 per million output tokens on its international deployment.

Both changes are 20 to 25 percent reductions against the model each one replaces. For a solo operator or a small team that runs agents every day, the model bill is usually the largest variable cost, so the new price lists change the arithmetic of which model handles which job.

What changed at Anthropic

The reduction on Fable 5.2 is 20 percent on both input and output. Anthropic's pricing documentation lists the full current lineup per million tokens:

  • Claude Fable 5.2: $8 input, $40 output.
  • Claude Opus 5.5: $4 input, $20 output.
  • Claude Sonnet 5.5: $2 input, $10 output.
  • Claude Haiku 5.5: $0.10 input, $0.50 output for prompts up to 100,000 tokens.

Two modifiers on the same page matter for agent workloads. Cached input that is read again costs $0.20 per million tokens on Fable 5.2, which is one fortieth of the standard input price. Agents resend the same instructions and files on every step, so cache reads make up a large share of a long run. The Batch API applies a 50 percent discount to input and output for work that does not need an immediate answer.

Anthropic also published its own benchmark table with the release. On FrontierSWE v2, an agentic coding test, the company reports 69.9 percent for Fable 5.2 against 62.3 percent for Opus 5.5. On AutomationBench, which the table labels as business workflows, it reports 51.4 percent for Fable 5.2 and 42.5 percent for Opus 5.5. The same table shows Google's Gemini 4 Carbon ahead of Fable 5.2 on AutomationBench at 53.9 percent. These are vendor-reported figures.

What changed at Alibaba

Alibaba's price list shows Qwen4-Max at $1.50 and $4.50 per million tokens for requests of up to one million tokens. The previous flagship, Qwen3.8-Max, is listed on the same page at $2 and $6. That is a 25 percent reduction on both sides.

The rollout came in two stages. AlphaSignal reported that the preview was reachable through Alibaba's Token Plan subscription, that the plan works with clients that use OpenAI-compatible or Anthropic-compatible protocols, including Cursor, Cline and Codex, and that Alibaba had not published benchmarks or a model card at that point. Crypto Briefing reported the full launch on November 16, with a preview window that runs until November 25.

According to Crypto Briefing, Alibaba says Qwen4-Max scored 75.7 on the DeepSWE v1.1 coding benchmark, uses about 50 percent fewer tokens than Qwen3.8-Max, and lowers average task cost by more than 60 percent against that model. The same report notes that independent rankings place Qwen4-Max behind some competitors on performance while crediting its cost per task. Alibaba plans smaller open-weight models in the Qwen4 series, and the flagship stays closed.

The same run at five price points

The following figures are arithmetic on the list prices above, not a measured workload. They price one agent run that reads 2 million input tokens and writes 200,000 output tokens, with no caching and no batch discount.

  • Claude Fable 5.1: $30.00.
  • Claude Fable 5.2: $24.00.
  • Claude Opus 5.5: $12.00.
  • Claude Sonnet 5.5: $6.00.
  • Qwen4-Max: $3.90.

The spread between the top and bottom of that list is larger than either price cut. A 20 percent reduction moves the most expensive option from $30 to $24. Choosing a different tier for the same job moves it by a factor of two to six.

The comparison has a limit that the vendors themselves point to. Models do not use the same number of tokens for the same task. Alibaba's claim about Qwen4-Max is that it needs about half the tokens of its predecessor, which would lower the bill more than the price change does. A model that fails a task and has to be run again costs more than its list price suggests. The site's guide to measuring cost per successful agent workflow explains why the completed task, not the token, is the unit to compare.

What this means for small operators

Earlier coverage on this site of routing work between cheap and frontier models described the practice of sending routine steps to a low-cost model and reserving the expensive one for hard steps. The new price lists do not change that practice. They widen the set of jobs where the top tier is affordable, and they add a lower-priced flagship option for operators willing to test a second vendor.

Protocol compatibility lowers the cost of that test. AlphaSignal notes that a supported client can be pointed at Qwen4-Max by changing the provider address and API key, and that feature parity can still vary for tool schemas, streaming and token accounting. A team that runs several agents in parallel multiplies whatever per-run difference it finds.

Practical steps

The following steps are general guidance drawn from the sources above.

Reprice the current workload first. An operator already on Fable 5.1 gets the 20 percent reduction by moving to Fable 5.2. Anthropic lists both models, so the change is not automatic.

Check cache and batch use before switching vendors. Cache reads at one fortieth of the input price and a 50 percent batch discount can lower a bill more than a move to a cheaper model.

Test a second model on real tasks. Vendor benchmark tables use different tests and cannot be compared with each other. The site's guide to evaluating an agent workflow covers how to run the same tasks on two models and compare completion and cost.

Record tokens per completed task. Price per million tokens is one input. Tokens used and the share of runs that finish correctly decide the actual cost. Readers new to the category can start with the agent fundamentals.

What to watch

Crypto Briefing reports that the Qwen4-Max preview window closes on November 25 and that Alibaba plans to release smaller open-weight Qwen4 models. Alibaba has not given dates for those releases in the sources reviewed here. Independent evaluations of both new models are still arriving, and they will show whether the vendor-reported results hold on tasks outside each company's own tests.

Sources