feat(deepseek): DeepSeek-V4.1-Flash, and V4 Pro's retirement onto it

[agentforce314]

What

deepseek-flash — DeepSeek-V4.1-Flash — is DeepSeek's current line: 1M context, 384K max output, thinking on by default, and the first DeepSeek model that accepts an image. It folds in the retired deepseek-v4-flash-vision-exp, so a pasted screenshot no longer needs a fusion model on this provider.

Spec and rates read from api-docs.deepseek.com/quick_start/pricing on 2026-09-10.

The DeepSeek line collapses onto it

It becomes the provider default and the whole subagent tier table (opus/sonnet/haiku). That's the vendor's arrangement rather than a simplification — DeepSeek says V4.1 Flash "has comprehensively surpassed V4 Pro in performance, cost, speed, and total time" and from 2026-09-14 routes every deepseek-v4-pro request to it at the Flash price. An opus-tier row naming v4-pro buys nothing after that date and costs up to 4.4× more before it.

deepseek-v4-pro, deepseek-v4-flash, deepseek-chat and deepseek-reasoner all still resolve and stay listed, so a pinned session keeps working (and the subagent resolver doesn't degrade it to inherit).

Pricing moves on two axes, both read off the request's timestamp

	input	output	cache hit
V4.1 Flash (off-peak)	$0.15	$0.60	$0.003
V4 flash card it replaces	$0.22	$0.66	$0.007
V4 Pro (off-peak, until cutover)	$0.66	$1.98	$0.022

• The flash re-card is cheaper, and the retired flash ids share it — DeepSeek serves them from V4.1 Flash and bills them at its price. Leaving them on the V4 card would over-report a legacy-id session 1.5× on input and 2.3× on cache read.
• deepseek-v4-pro keeps its own card until 2026-09-14 12:00 Beijing (04:00 UTC) and prices as Flash from that instant. A date axis rather than a card edit because both sides are live: flattening pro onto the flash card outright would restate an August session 4.4× low, and leaving it on the pro card would over-report every later request by the same factor. compute_cost already carried the request timestamp for the peak schedule, so the retirement rides the same value.
• The peak/off-peak schedule composes with both, unchanged.

On the agent token mix issue #904 measured (95.6% cache read), 1M tokens off-peak now costs $0.0107 against $0.0536 on the old pro card.

One structural change

ModelConfig.claims_prefix. deepseek-flash's prefix base under get_model_config's fallback (key.rsplit("-", 1)[0]) is the bare family name, so registering the row would have captured deepseek-chat, deepseek-reasoner and OpenRouter's deepseek/… ids — the last of which test_deepseek_prefix_cache.py pins at the 200K default as "decision #1". The row opts out of the fallback and stays reachable by its exact id. Every other row keeps the default True.

Tests

New coverage for the cutover (instant, Beijing→UTC conversion, composition with the peak windows, and that history isn't restated), the retired ids sharing the flash card, the prefix opt-out, and vision being flash-only. Existing DeepSeek cards, windows and subagent-tier expectations updated.

537 passed  # deepseek/pricing/provider/subagent suites
1184 passed, 4 failed  # every deepseek-touching test; all 4 failures reproduce on main

🤖 Generated with Claude Code