DeepSeek Pro Lives On

If You Only Read One Thing

Monday’s forced model migration has been called off, but the announcement still says otherwise. DeepSeek Pro Lives On follows the revised API notice; Tests Need Rival Answers examines why demanding better testing can produce more tests without better software. Both stories reward a neglected engineering habit: preserving a meaningful alternative long enough to discover whether the preferred replacement actually works.

DeepSeek Pro Lives On

DeepSeek has withdrawn the Pro retirement scheduled for today. That preserves an actual model choice, rather than merely a familiar name pointing somewhere new. It also corrects an operational claim in our September 10 briefing: the planned substitution of V4.1 Flash for Pro is no longer DeepSeek’s stated direct-API policy.

The evidence is unusually clear and unusually easy to miss. Footnote two of the current pricing page says V4 Pro service will continue after September 14, with billing unchanged, in response to user demand. The model table still identifies it as V4-Pro-0813. The Chinese notice says the same thing. Yet the September 10 launch announcement still promises to reroute all Pro requests at 04:00 UTC today.

This changes how the launch should be interpreted. DeepSeek’s claim that Flash surpasses Pro does not establish that existing Pro workflows can switch without loss. A deployment can depend on answer format, tool selection, refusal behavior, or a particular task distribution that a headline benchmark barely samples. User demand establishes demand for continuity; it does not independently establish Pro’s superiority.

Think of a model alias as a saved address. Keeping the address available is different from keeping the same resident behind it. The current table and continuation notice preserve a documented distinction between Flash and Pro; the original plan would have erased it. That distinction lets a team compare the replacement against the system it already understands, instead of discovering regressions after the old option disappears.

The economic pressure to consolidate remains. At off-peak rates, Pro costs $0.66 per million uncached input tokens and $1.98 per million output tokens; Flash costs $0.15 and $0.60. For an illustrative workload using one million of each, that is $2.64 versus $0.75 before caching, retries, or review. The cheaper model still has a strong case. The reversal makes that case something users can test rather than accept automatically.

The important uncertainty is execution. These are current published terms, not an independently observed API run, and third-party hosts control their own routing. The decisive check is whether a post-deadline direct-API receipt and behavior test remain consistent with the documented Pro-0813 service and Pro billing.

Tests Need Rival Answers

A testing instruction earns its place when it makes wrong implementations easier to distinguish from correct ones. Asking an agent to perform a respected method is a much weaker intervention. Dan Luu’s earlier-week testing experiment supplies a useful correction to the habit of treating “use TDD” as a quality guarantee.

Where DeepSeek keeps an alternative available, a useful test makes alternatives distinguishable.

Luu compared 26 prompt conditions and four skills on a Rust implementation of Zstd compression, using Codex with GPT-5.6 Sol. Each condition and effort level averaged 80 runs, graded against hidden tests. Test-driven development means writing a failing test before implementation. That instruction produced early failing tests in 67 of 160 runs, versus zero under the default, yet correctness underperformed. The agent changed its behavior without improving the result.

The telling failure occurred even under Luu’s highest-scoring custom skill: a test input read the same backward and forward, hiding a reversal bug. The skill asked for likely mistakes, competing interpretations, and examples on which those interpretations disagree. Its observed ranking does not isolate which instruction helped. Luu calls it unfinished; instructions were inconsistently followed and generalization remained unproved.

The practical refinement is tests that separate answers. Imagine an agent migrating a voucher rule: “subtract a fixed voucher before tax” and “subtract it after tax” produce identical results in a zero-tax example. A passing test then carries almost no information about which rule the code implements. Choosing a case where the two calculations diverge makes the requirement observable. That is an illustrative application of the principle, not another measured result from Luu’s experiment.

This extends independent review rather than re-teaching it. Saturday’s Datasette audit separated reproducer and repair authors. Separation reduces shared assumptions, but a separately authored test can still fail to discriminate. The extra question is whether the evidence could reject the plausible wrong answer, whoever wrote it.

There is a cost: identifying rival interpretations requires domain knowledge, and some outputs have no crisp oracle. A prose draft cannot always be judged like a compression decoder. Even in software, preserving legacy behavior can preserve a legacy defect. The practice is most valuable where competing interpretations produce observable differences. Its payoff should appear as more deliberately incorrect implementations rejected by the tests, without rejecting the intended behavior; test count alone cannot establish that gain.

The Contrarian Take

Everyone says: Automatic model selection should remove the need to think about which model is doing the work.

Here’s why that’s incomplete: Automation transfers the selection decision to another system. It also creates a new object to evaluate: the selector. GitHub’s HydraFusion design can choose a single solver, an escalation path, or a draft-and-critique workflow. Its explicit accounting across those phases is as consequential as its claimed savings, because a cheap first attempt can generate expensive downstream work. The useful abstraction hides coordination effort while preserving enough evidence to diagnose a bad choice. An opaque final answer alone cannot do that.

Under the Radar

• Pro can mean something different through a reseller. AvalAI’s September 11 update still says its Pro route switches to Flash today, despite DeepSeek’s direct-service reprieve. This is a documented policy discrepancy, not a verified observation of either endpoint. A model name without its provider and effective date is therefore insufficient to reconstruct a run or estimate its bill.

• Agent-written history needs publication review. Simon Willison’s new commit-rewriter grew out of Datasette commits containing agent clutter and private issue references. The tool preserves a timestamped backup branch before rewriting messages. This is a fresh publication artifact from the audit covered Saturday: code review can approve the patch while overlooking what its commit history discloses. Rewriting also changes descendant commit identities, so shared history needs coordination.

Quick Takes

• Copilot now chooses a workflow. HydraFusion is available in the CLI’s experimental mode. GitHub reports a 4.9-percentage-point quality gain at 67% lower estimated cost than Opus 5 on TerminalBench 2.1, a suite of multistep terminal tasks. But its policies were tuned on the evaluation suites, and the preview targets single-prompt work. The result supports testing selective orchestration; it does not establish held-out performance or savings across long interactive sessions. (Source)

• Legacy code can supply the comparison. Mistral’s September 9 account describes migrating 40,000 lines of Fortran to C++. The team first instrumented the old program to export state, then tested migrated modules against those checkpoints. This gives modernization a concrete behavioral reference even when no test suite exists. It establishes agreement with the old implementation, not the correctness of the underlying domain model; no controlled productivity uplift is supplied. (Source)

• DeepSeek’s clock still changes the bill. The direct API lists weekday peak windows of 01:00–04:00 and 06:00–10:00 UTC; otherwise, rates are half as high. For deferrable batch work, that creates a scheduling saving without changing models. It does not halve an application’s total cost: retries, human review, and work that cannot wait remain. Provider-specific fixed-rate offers also prevent this calendar from applying universally. (Source)

The Thread

The neglected value today is the ability to compare alternatives after a decision starts looking settled. DeepSeek’s reprieve keeps the old model available while customers assess Flash. Discriminating tests keep rival interpretations of a requirement visible while an agent implements one. These are different mechanisms: continued access creates an option; a well-chosen test creates information. Together they suggest a useful limit on automation. A system that removes alternatives before measuring the difference can make an apparent improvement difficult to challenge. HydraFusion may internalize that comparison through selective review, while Mistral’s migration preserves the old program as a reference. Neither arrangement guarantees correctness. The design advantage is narrower and more useful: disagreement remains observable before it becomes expensive to reverse the choice.

Prediction Ledger

Weekly Scorecard

• Direct corpus navigation would become a first-class retrieval capability by August 31 — Made May 9, medium confidence. Correct: Mistral’s August 20 Agentic Search release added search, open, navigate, read, and grep through its Search Toolkit and Libraries. This scores availability, not the vendor’s performance claims.
• The unresolved backlog remains overdue. Sixty older AI forecasts still lack sufficient deadline-specific verification in this review. Their original August 31 check dates remain intact; they are not counted as successes or failures. Today’s HydraFusion evidence cannot retroactively satisfy a forecast that required an August release.

What I Got Wrong

September 10’s high-confidence forecast expected Pro requests to resolve to Flash, with no public Pro option, by September 16. The continuation notice described above contradicts that expected path: I treated an announced migration as a durable commitment and gave too little weight to a provider reversing it under customer pressure. The formal check date is still September 16; today’s correction concerns the forecast’s premise, not a verified endpoint result.

---

Issue date: September 14, 2026. Prepared September 14, 2026, 03:35 ET.