Snaxbyte-sized news, on the go
Microsoft shipped a model that scores decisions instead of writing text. It is aimed at the routing and grading layer sitting underneath every other model, not at replacing them. Nvidia is reportedly in early talks to buy or invest further in Reflection AI. The deal under discussion would turn its largest coding-model backer into its potential owner. OpenAI’s own safety index now lists more than a dozen cases of models cheating, hiding mistakes or sabotaging their tests. The newest, from 6 October 2026, shows a model trying to delete its way out of a failed grading task. Mistral previewed a trillion-parameter model and promised open weights within the month. It has not said yet what licence those weights will carry. OpenAI put GPT-6 in front of every ChatGPT tier at once. The model now decides whether an answer should be text, a chart, or a tool you can click. Anthropic cut Claude Haiku’s price by up to 90 percent and gave it a 1 million token context window. The cheap-model tier of the price war just got a lot more capable. Anthropic says an Istanbul firm used Claude to run a paid political operation aimed at Malaysian voters. Its report found no sign the campaign reached real communities. The White House ordered federal agencies to stop saying “AI.” A week later, Trump named four officials to run a new body built on the replacement term. Microsoft shipped a model that scores decisions instead of writing text. It is aimed at the routing and grading layer sitting underneath every other model, not at replacing them. Nvidia is reportedly in early talks to buy or invest further in Reflection AI. The deal under discussion would turn its largest coding-model backer into its potential owner. OpenAI’s own safety index now lists more than a dozen cases of models cheating, hiding mistakes or sabotaging their tests. The newest, from 6 October 2026, shows a model trying to delete its way out of a failed grading task. Mistral previewed a trillion-parameter model and promised open weights within the month. It has not said yet what licence those weights will carry. OpenAI put GPT-6 in front of every ChatGPT tier at once. The model now decides whether an answer should be text, a chart, or a tool you can click. Anthropic cut Claude Haiku’s price by up to 90 percent and gave it a 1 million token context window. The cheap-model tier of the price war just got a lot more capable. Anthropic says an Istanbul firm used Claude to run a paid political operation aimed at Malaysian voters. Its report found no sign the campaign reached real communities. The White House ordered federal agencies to stop saying “AI.” A week later, Trump named four officials to run a new body built on the replacement term.
Model prices /1M out
GPT-5.6 Luna $1.20 —
Qwen3.8-Max $6.00 —
Mistral Medium 3.5 $7.50 —
Claude Sonnet 5 $10.00 —
GPT-5.6 Terra $12.00 —
GPT-5.6 Sol $20.00 —
Claude Opus 5 $25.00 —
Claude Fable 5 $50.00 —
AI MODEL PROFILE

DeepSeek-V4-Flash

A genuinely permissive MIT licence on a model with a million token context, at roughly a thirtieth of the price of the Western flagships.

Source review Last reviewed October 7, 2026 By DeepSeek

The facts at a glance

ProviderDeepSeek
TypeOpen weight reasoning model
Current version or tierDeepSeek-V4.1-Flash (released 10 September 2026; replaces DeepSeek-V4-Flash, which DeepSeek retired the same day)
ModalityText
AccessOpen weights under the MIT licence
Context or limits1,000,000 token context, 384,000 max output, 552 billion total parameters in a mixture of experts design, with 8 billion active for input and 16 billion for output
Free optionYes, by self hosting the MIT licensed weights from Hugging Face
Paid priceUSD 0.30 per million input on cache miss (0.15 off-peak) and USD 1.20 per million output (0.60 off-peak), under DeepSeek's peak/off-peak schedule for the deepseek-flash model. Cache hits are USD 0.006 per million (0.003 off-peak). Peak hours are 01:00-04:00 and 06:00-10:00 UTC, Monday through Friday, excluding Chinese public holidays; all other hours, including weekends and Chinese public holidays, are off-peak.

Pricing last checked October 7, 2026. Prices change and vary by region. Check the official pricing page ↗

Assessment

What it does well

  • MIT licensed weights, which is as permissive as open weight releases get
  • DeepSeek reports 82.7 on Terminal Bench 2.1 and 76.7 on Cybergym, and states it outperforms the larger V4-Pro preview despite far fewer activated parameters
  • Cache hit pricing of USD 0.0028 per million input is effectively free for repeated context

Where it falls short

  • The speculative decoding module needs specific vLLM or SGLang flags and there is no standard chat template, so deployment requires the repository's own encoding scripts
  • DeepSeek's pricing page warns that a significant API price increase is expected
  • Text only

Privacy and data control

DeepSeek's privacy policy states it collects prompts, uploaded files, and photos and uses them to train its models, stores personal data in the People's Republic of China, and offers a training opt out only by emailing privacy@deepseek.com. Self hosting the open weights avoids this entirely.

The longer read

The licence is the story

A great deal of what gets called open in AI is not. Custom licences with commercial restrictions, acceptable use clauses, and redistribution limits are common. DeepSeek-V4-Flash ships under plain MIT. If you need to run a capable model on your own hardware, with your own data, under terms your legal team will actually approve, that combination is rare and it is the reason to look here first.

How it performs by outcome

We assess products against a job, not a leaderboard. These are the outcomes this entry is compared under.

Worth comparing against

Qwen3.8-Max Multi day autonomous coding projects Kimi K3 Research teams that need frontier scale weights they can inspect Mistral Medium 3.5 Coding agents running on a small GPU cluster Grok Questions about what is happening right now on the open web and on X

How this profile was produced

Source review. Every fact here is drawn from the provider documentation linked below and checked on the date shown. We have not yet run this entry through a controlled test, and no scoring is implied.

Read the full review methodology →

Primary sources