Fireworks puts DeepSeek V4.1-Flash level with Astra on coding; test it yourself

*A four-way tie inside one point on a hosting vendor's own run is a reason to run your pull requests through it, not a reason to switch*

Opening read

A model that runs at a fifteenth of Astra's price now scores level with it on coding, on the table of a vendor that sells the cheap one. The right response is to run the comparison on your own code this week rather than to believe or dismiss the number. The savings are real enough to justify the test, and the evidence is thin enough that the test is the only thing that settles it.

Open models: the fallback just claimed the frontier

DeepSeek released v4.1-Flash last week and retired its V4 Pro in favor of it, and this brief read the release as a cheaper fallback worth keeping warm Latent Space (https://www.latent.space/p/ainews-deepseek-v41-flash-763b-p8b). Fireworks, which hosts the model, now says the fallback is the frontier. On DeepSWE, a benchmark of real software fixes, v4.1-Flash solved 74.34% of tasks on the first try against 74.12% for GPT-6 Astra, and Terminal-Bench 2.1, a test of command-line work, has Astra a point ahead at 87.5% to 86.5% Fireworks AI (https://fireworks.ai/blog/DeepSeek-V4.1-Flash-Astra). The obvious read is to move coding agents to the open model and pocket the difference, and a smart reader could believe it on those numbers.

Vendor benchmarks describe the vendor's environment, and this one has two problems. Fireworks profits from the model it is ranking, and the gap across its top four, from Claude Opus 5 at 73.65% up to DeepSeek, is under one point Fireworks AI (https://fireworks.ai/blog/DeepSeek-V4.1-Flash-Astra). I do not know the run-to-run variance and Fireworks did not publish it, so I read that as a tie. The record on DeepSeek's July model is mixed, and it splits by who ran the test and against what. In August, Together AI's run put GPT-5.6 Luna 14 points ahead of DeepSeek's July release on 113 DeepSWE tasks, while Snowflake's run the same month had that same July model ahead of the leading proprietary model it tested on ADE-bench, a test of data-team tasks, at 74.4%, and Interconnects judged the July update ahead of Luna on the Pareto frontier, meaning the best available trade of capability against cost Together AI Blog (https://www.together.ai/blog/deepseek-v4-flash-0731-vs-gpt-5-6-luna-on-deepswe-cost-and-coding) Snowflake Blog (https://www.snowflake.com/en/blog/dynamic-model-routing-open-models-cortex-ai/) Interconnects (https://www.interconnects.ai/p/latest-open-artifacts-23-laguna-s21). Those runs used different benchmarks and different comparison models, so the answer moves with the test at least as much as with the host. None of them tells you what happens on your code.

The cost claim needs the same care. Fireworks prices its run at $0.430 per task, but the model read 36.9 million input tokens (the chunks of text it takes in) per task to get there Fireworks AI (https://fireworks.ai/blog/DeepSeek-V4.1-Flash-Astra). At that volume, cost tracks prompt caching, meaning the host reuses text the model has already read so it is not billed twice, and the host's price sheet more than it tracks the weights, the downloadable model file, so the one-fifteenth-of-Astra headline is a Fireworks number. For an engineer deciding where coding tasks route, that is the live question: a cheap path exists, and its price is host-specific. For a business leader, the coding-agent line item just became negotiable.

The test is simple. Take 20 recent merged pull requests, meaning code changes already accepted into the main codebase, that have passing tests, revert each one, and run both models through the same agent harness, the software that lets a model edit and run code. Count how many make the existing tests pass again, and pull spend and input tokens per task from the bills. If DeepSeek's pass count sits inside your review tolerance, route routine fixes to it and keep Astra for what it fails; if input tokens per task land in the tens of millions, price caching before you commit. Do not switch on Fireworks' table alone. One more limit: the US-hosted endpoints for regulated industries are promised, not shipped, so a bank cannot run this test on the intended production path yet.

Pacing the frontier: Nvidia's no is the tell

Yesterday I sorted Dario Amodei's plan to slow the most capable models by who could deliver each part; today the part he does not control got its answer. His three pieces are outside evaluators inside the labs, coordination among democratic countries, and international cooperation, and so far Amodei committed his own lab to the first, OpenAI followed, and Musk and Hassabis endorsed the whole Nate Jones (https://natesnewsletter.substack.com/p/us-china-ai-human-flourishing) Don't Worry About the Vase (https://thezvi.substack.com/p/we-must-pace-the-frontier). The rest runs on chips. Jensen Huang told Trump a slowdown is not going to happen, a direct split from Amodei, Musk, and Altman TechCrunch AI (https://techcrunch.com/2026/09/14/nvidia-ceo-jensen-huang-tells-trump-were-not-going-to-let-an-ai-slowdown-happen/). A business leader budgeting around a calmer release calendar should stop; the evaluator commitments are the checkable deliverable.

AI video: China cleared prime time in six months

Mango Excellent Media's adaptation of The Later Journey to the West aired August 31, the first series made entirely by AI to land in prime time on a Chinese provincial satellite channel; every frame came from ByteDance's Seedance 2.0 and 2.5, and production took six months against one to two years for a filmed drama. Investors added 13.3 billion RMB to Mango's value in five days, but the volume story is older: more than 74 percent of the 367,000 microdramas released on Chinese platforms in the first half of 2026 were already AI-generated ChinaAI (https://chinai.substack.com/p/chinai-374-chinas-first-ai-generated). The stock move is not audiences approving; it prices a cost cut the short-form market had already accepted. Seedance 2.5 makes clips of 4 to 30 seconds with audio in the same pass, so an episode is a stitching job OpenRouter Blog (https://openrouter.ai/blog/insights/seedance-2-5-review/). Stitching is cheap to try. A media or marketing leader should pilot generated video where speed beats polish, because the market has already decided the economics work.

Private data: Snowflake tunnels out instead of in

Snowflake's new Data Connectivity Proxy lets Openflow, its managed service for moving data in, reach on-premises databases and private Kafka brokers (message-streaming systems) through an outbound encrypted connection on port 443, so the customer opens no inbound firewall rule Snowflake Blog (https://www.snowflake.com/en/blog/data-connectivity-proxy-openflow/). That fits the deny-all-inbound shops regulation creates, and credentials stay in Snowflake while the agent passes bytes without reading them. The catch for an IT leader is that routing the agent through an enterprise forward proxy is not supported, so if policy forces all outbound traffic through one, this is a security-policy conversation before it is a pipeline.

Augur calls

• Call: Fireworks publicly lists a US-hosted DeepSeek-V4.1-Flash endpoint on or before 2026-11-30.
• Criteria: A Fireworks product page, documentation page, or blog post naming a US-region DeepSeek-V4.1-Flash endpoint counts as a hit; no such public listing by the due date is a miss.
• Due: 2026-11-30
• Odds: 70%
• Call: Snowflake announces DeepSeek-V4.1-Flash as an available model in Cortex AI on or before 2026-12-31.
• Criteria: A Snowflake blog post, release note, or documentation page listing DeepSeek-V4.1-Flash in Cortex AI counts as a hit; absence by the due date is a miss.
• Due: 2026-12-31
• Odds: 55%

What to watch

• Fireworks has promised US-hosted endpoints and training support for DeepSeek-V4.1-Flash, and the shipping date decides whether regulated buyers can run the test above. Fireworks AI (https://fireworks.ai/blog/DeepSeek-V4.1-Flash-Astra)
• Together AI's next DeepSWE comparison would show whether the August 14-point gap has closed on a second host's run. Together AI Blog (https://www.together.ai/blog/deepseek-v4-flash-0731-vs-gpt-5-6-luna-on-deepswe-cost-and-coding)
• Nvidia's public stance against a slowdown sets up a collision with Amodei's chip-control proposals, and the White House response is the signal. TechCrunch AI (https://techcrunch.com/2026/09/14/nvidia-ceo-jensen-huang-tells-trump-were-not-going-to-let-an-ai-slowdown-happen/) TechCrunch AI (https://techcrunch.com/2026/09/12/anthropic-ceo-outlines-plan-to-pace-the-frontier/)
• Mango Excellent Media's next commission will show whether one prime-time series was a stunt or a production model. ChinaAI (https://chinai.substack.com/p/chinai-374-chinas-first-ai-generated)

Verification notes

Every figure in the Astra comparison comes from Fireworks' own run, with no error bars, from a vendor that sells hosting for the model it ranked first, and its US-hosted endpoints and training support remain promises Fireworks AI (https://fireworks.ai/blog/DeepSeek-V4.1-Flash-Astra). The cross-host comparison leans on two other vendors' runs, Together AI's and Snowflake's, on different benchmarks against different comparison models, plus one newsletter's judgment, so it shows variance rather than a ranking. The Mango series details, its five-day valuation move, and the microdrama volume figure rest on a single newsletter ChinaAI (https://chinai.substack.com/p/chinai-374-chinas-first-ai-generated).

Inspect the evidence behind this dispatch. (https://augurdispatch.ai/evidence/2026-09-15?utm_source=augur&utm_medium=web&utm_campaign=dispatch-2026-09-15)