Fireworks has a much cheaper coding model now lands in Astra's accuracy band. Augur's recommendation is to test the saving on your own code before changing the default. The vendor's price is a reason to run a comparison, not a promise, about your next bill. This is the Augur Dispatch for Tuesday, September 15th, 2026. A quick disclosure before we begin. Portions of the Augur Dispatch Podcast are generated using artificial intelligence. The written dispatch contains the full source trail and verification notes. On Fireworks Software Benchmark, DeepSeek V 4.1 Flash solved about 74% of tasks on the first attempt, essentially level with GPT-6 Astro. The reported cost was 43 cents per task against roughly $6.52 for Astro. Imagine an engineering manager looking at that difference while the team has a queue of routine fixes. It would be strange not to investigate. It would also be expensive to assume every item in the queue looks like the vendor's test. The purchasing decision is whether some of that work can move to a cheaper model without moving the cost into somebody else's afternoon. Fireworks gives us more reason to call this an accuracy band than the headline alone suggests. It reports variation between runs of roughly one and a half to three percentage points. The spread across its top four models is smaller than that. That doesn't prove the models are equivalent everywhere. It means the small differences on this table aren't a sound basis for declaring a winner, even before we account for the vendor's commercial interest in selling Deep Seek hosting. A second test, focused on command line work, puts Astra about one point ahead. Again, this is Fireworks' evaluation, not an independent result from your engineering team. The older DeepSeq model also gives us a useful warning. Together, AI tested the July release on 113 software tasks and found Luna ahead by about 14 percentage points. Snowflake tested that July model on a different benchmark for data team work and found it ahead of the leading proprietary model in its comparison. Those results aren't votes in the same election. Different tasks and comparison models can produce different answers, and neither August test settles what the new model will do today. Auger's read is that the broad opportunity deserves attention, while the decision still belongs at the level of the work you actually need completed. The cheap bill has an important detail tucked inside it. Fireworks Run read nearly 37 million input tokens per task. Tokens are the small chunks of text a model processes, including code and earlier tool results that an agent may read repeatedly. Prompt caching lets the host reuse previously processed material and charge less for those reads. It doesn't make them free. Fireworks reports that almost all the input was cached, yet those discounted reads still made up about 60% of the bill. Picture a finance lead hearing that an agent barely costs anything to run, then discovering that the estimate assumes a particular host's cash pricing and a very high reuse rate. The spreadsheet has acquired a footnote. Naturally, the footnote needs its own spreadsheet. The architecture helps make lower cost possible, but the price per completed task still depends on how the agent works and what the host charges for it. A model that rereads a lot of context can look cheap under one arrangement and less attractive under another. Before negotiating a larger commitment, ask for cached input, uncached input, output, and failed attempts to be visible in the bill. You can't manage a saving you can't explain. There is a manageable way to test this without turning the whole engineering team into a benchmark department. Have the engineering lead choose 20 recent accepted code changes with passing tests. In isolated copies of the code base, return each task to the point before its change, then ask both models to repair it through the same agent setup. That setup is the software that lets a model inspect files, edit code, and run checks. Keep the task instructions and attempt budget comparable, hide the original repair, and don't let either model rewrite the tests used to grade its work. Compare accepted repairs and total spend against Astra on those same tasks. Include the time a person spends reviewing the answer and the extra cost when a failed Deep Seek attempt goes to Astra. Agree on an acceptable quality gap and review time limit before the run, then review the results on September 22nd. Stop if the endpoint fails your data policy, if a model changes the grading tests, or if extra review wipes out the savings. A useful result earns a second sample. It doesn't earn automatic access to production. The advantage, if it holds, is a growing record of which tasks can use the cheaper route safely. Every competitor can buy model access. They don't have your task history and tests. For a bank or another buyer requiring fireworks promised US hosted endpoint, that specific path is still coming soon. An available endpoint elsewhere isn't permission to send regulated code there. Model budgets aren't the only plans depending on promises this week. The argument over slowing frontier development now has a clear public split. Dario Amade's proposal combines embedded outside evaluators, coordination among leading AI companies in democratic countries, and international cooperation. Anthropic has committed to the evaluator step, and OpenAI has followed, according to the reporting in today's dispatch. TechCrunch reports that Jensen Huang pulled President Trump a slowdown wouldn't happen. Trump publicly sided with him. That is a political position, not an enacted policy, but it isn't a basis for assuming the release calendar will get calmer. For the executive planning staffing or vendor reviews, Augur recommends watching whether evaluator access becomes concrete and whether governments enable coordination or change chip policy. A public endorsement doesn't tell you who can inspect a lab or what happens after an inspection finds a problem. The checkable commitments matter more to that planning decision than agreement on a conference stage. A finished television series offers a different kind of evidence. Something was made, although its profitability remains unresolved. Chen AI reports that Mango Excellent Media brought an entirely AI-generated adaptation of the later Journey to the West to Chinese Provincial Satellite Primetime on August 31st. Production took six months, compared with one to two years for a film drama. The same newsletter reports a sharp rise in Mango's market value and a large existing market for generated microdramas. Those figures don't establish that audiences like the series or that it made money. The production team framed it as a demonstration of feasibility rather than profitability. Consider a marketing producer asked to turn a product brief into a short video by Friday. Faster generation could help, but the producer still has to approve the result, fix continuity, and throw away the versions that don't work. The render button has not been promoted to Creative Director. OpenRouter says CDance 2.5 can generate clips up to 30 seconds with audio and the same pass. Augur recommends a small comparison only where speed matters and rights are cleared. One brief, owned by the production lead, against the current workflow, reviewed on September 22nd. Count everything through approval. Stop if continuity or brand standards fail, rights are unclear, or retries remove the saving. A second successful brief would justify expansion. A rising share price would it. For teams moving private data, Snowflake's new connection option could remove a different source of delay. Its data connectivity proxy lets a small program inside the customer's network open an encrypted outbound connection to Snowflake. The customer doesn't have to open an inbound firewall rule. Source credentials stay in Snowflake, and the program passes the traffic without inspecting or storing its contents. The limitation matters to the network engineer signing off the design. The program needs direct outbound access. Routing through a company-mandated forward proxy isn't supported. Check that policy before scheduling a migration. Augur's two dated calls remain unchanged. Fireworks publicly lists a US-hosted DeepSeek V4.1 flash endpoint by November 30, 2026 at 70% odds. Snowflake announces that model as available in Cortex AI by December 31st, 2026 at 55% odds. These are testable predictions with public listing criteria, not announcements of availability. The written dispatch carries the exact hit or miss rules, and the website tracks the calls so they can be judged later. Pilot the coding comparison or the data rules allow it. Watch the promised hosting path and the policy commitments. Keep the video test bounded by the cost of an approved result. The written brief with every source is at augurdispatch.ai. That is the Augur Dispatch for Tuesday, September 15th, 2026. Thank you for listening.