← AI Deflation Watch

Grok 4.6 and Qwen3.8 just moved the cost curve in two different ways.

One model got materially smarter without a list-price increase. The other put a 2.4-trillion-parameter MoE flagship with a million-token context window at the same headline price. Neither tells you what your task costs—but both change the set of models worth testing.

The cleanest deflation datapoint is Grok 4.6. SpaceXAI reports an Artificial Analysis Intelligence Index score of 61, up from 56 for Grok 4.5. Its API price starts at the same $2 per million input tokens and $6 per million output tokens charged for 4.5.

That is not a token-price cut. It is capability deflation: more intelligence offered at the old meter rate. If the improvement raises your first-pass success rate or removes retries, cost per successful task can fall even when the invoice line items do not.

Grok 4.6: the vertical move

SpaceXAI released Grok 4.6 on August 12 with a focus on long-running agents and interactive work. The company reports 61 on the AA composite, 69.9% on CursorBench 3.2, and 65.9% on DeepSWE 1.1. It also reports a longer supplemental training run and refreshed post-training trajectories. Those figures are useful because 4.5 is shown in the same launch table, but they remain benchmark evidence—not a guarantee for a specific workflow. (SpaceXAI announcement)

The API keeps a 500k-token context window and charges $2/$6, with cached input at $0.50 per million tokens. Requests above 200k context can use different rates, so the headline price is not the bill for a repository, research loop, or multi-hour agent run. (Grok 4.6 API documentation)

Qwen3.8-Max: the horizontal pressure

Qwen3.8-Max is a different signal. QwenCloud describes it as a 2.4-trillion-parameter mixture-of-experts model, and lists a second endpoint named qwen3.8-2.4t-a95b. It accepts text, images, and video; supports a 1M-token context window; and allows up to 131k output tokens. (Qwen3.8-Max model page)

Its standard API price is also $2/$6. The sharper datapoint is repeated context: implicit cached input is $0.25 per million tokens and explicit cache reads are $0.17. For workflows that resend a stable codebase, document set, policy library, or tool schema, the cache design can matter more than the uncached sticker price.

Qwen positions 3.8-Max as its strongest reasoning model and maps it to the highest-capability tier, but its current product documentation does not cite an independently measured task-cost result. That distinction is important: the architecture and prices are facts; the task economics still need to be measured. (QwenCloud model guide)

Published API facts at launch
DatapointGrok 4.6Qwen3.8-Max
ReleaseAug 12, 2026Current flagship
API list price$2 input / $6 output$2 input / $6 output
Cached input$0.50$0.25 implicit / $0.17 read
Context500k1M
Measured AA score61Not cited in current product docs
Architecture disclosureNot disclosed2.4T MoE; A95B endpoint

The metric that belongs on your curve

Cost per million tokens is an input. For an operating decision, the more useful unit is cost per successful task: model calls, retries, tool fees, latency, and human review divided by outcomes that clear a fixed quality bar.

EXPECTED TASK COST(calls + retries + tools + review) ÷ accepted outcomes

Grok 4.6 may lower that number by completing more work per attempt. Qwen3.8-Max may lower it through model competition, cache economics, or a better fit for multimodal and long-context tasks. The only defensible way to know is to run both against the same task set, rubric, harness, and reviewer policy—then repeat the assessment as the frontier moves.