Editorial Team
5 min read
DeepSeek’s announced July 24 retirement of deepseek-chat and deepseek-reasoner turns a formerly implicit routing choice into an explicit selection of V4 tier and thinking mode. The documented like-for-like destination is V4-Flash—not automatically V4-Pro—and its pricing differs sharply depending on cache use.
deepseek-chat and deepseek-reasoner; that announced deadline has now passed.deepseek-v4-flash: chat to non-thinking mode and reasoner to thinking mode.deepseek-v4-pro is a separate performance-and-price decision, not the documented compatibility replacement.DeepSeek, the provider of the V4 API models, told customers to replace its legacy deepseek-chat and deepseek-reasoner names before an announced July 24, 2026, 15:59 UTC cutoff. Its V4 release notice says the names would become inaccessible after that time. The consequential part is not the spelling change: the retired labels had been a compatibility layer over an underlying model and its two response modes.
The direct compatibility mapping is deepseek-v4-flash, with thinking explicitly disabled for the old chat behavior and enabled for the old reasoner behavior. DeepSeek’s model documentation identifies thinking as the default mode for both V4 tiers. That means a codebase that merely replaces deepseek-chat with deepseek-v4-flash can change behavior unless it also specifies non-thinking mode.
DeepSeek’s official V4 release notice gives the legacy aliases’ retirement time and V4-Flash compatibility routing. Source: DeepSeek API Docs.
V4-Flash and V4-Pro remain available through DeepSeek’s OpenAI-compatible Chat Completions and Anthropic-compatible interfaces. The V4 release announcement says the base URL remains unchanged; callers select the tier through the model parameter.
The aliases were deliberately designed to hide that selection. DeepSeek’s change log shows deepseek-chat had pointed to V3 in December 2024, while deepseek-reasoner was introduced for R1 in January 2025. By August 2025, the company had made them the non-thinking and thinking modes of one V3.1 model. They later moved together through V3.1-Terminus, V3.2-Exp and V3.2 before the April 2026 V4 transition.
That record limits the broader claim that an explicit identifier necessarily freezes a deployment: DeepSeek has changed the targets of its generic aliases before, but its current V4 documentation names two explicit model choices. What the deadline unquestionably removes is the provider-managed bridge from those legacy strings to V4-Flash’s two modes.
During the V4 transition, the old labels selected Flash and its response mode for the caller. The documented mapping now makes that selection visible and leaves any move to Pro as a separate choice:
| Former identifier | Transition destination | Intended mode | What it does not establish |
|---|---|---|---|
deepseek-chat | deepseek-v4-flash | Non-thinking | That Pro is needed |
deepseek-reasoner | deepseek-v4-flash | Thinking | That Flash and Pro perform identically |
V4-Flash is DeepSeek’s smaller, economy-oriented model, with 284 billion total parameters and 13 billion active parameters, according to the company’s release announcement. V4-Pro is the larger tier, listed at 1.6 trillion total parameters and 49 billion active parameters. Both are advertised as supporting thinking and non-thinking modes, a one-million-token context length, a 384,000-token maximum output, JSON output and tool calls.
DeepSeek describes Pro’s performance in promotional terms, including claims about agentic coding and reasoning benchmarks. Those are company claims, not a published assurance that every workload formerly sent to deepseek-reasoner needs Pro. The compatibility mapping instead says the starting point for that workload is Flash in thinking mode; a move to Pro should follow workload-specific evaluation.
The listed direct-API prices are per one million tokens. Cache-hit input is materially cheaper than cache-miss input, so a single input-price comparison can misstate the cost difference a deployment will see.
| Listed price per 1M tokens | V4-Flash | V4-Pro | Pro relative to Flash |
|---|---|---|---|
| Input, cache hit | $0.0028 | $0.003625 | 1.29× |
| Input, cache miss | $0.14 | $0.435 | 3.11× |
| Output | $0.28 | $0.87 | 3.11× |
| Concurrency limit | 2,500 | 500 | — |
These are DeepSeek’s published list rates and limits, not a forecast of any application’s bill. The ratios are simple price divisions within the same row. The documentation also says prices may change.
Thinking matters to this comparison because it changes the requested response mode, while output tokens have their own listed rate. A retained migration analysis warns that a mechanical chat-to-Flash rename leaves thinking enabled by default, potentially changing output-token use and latency. That is a plausible migration risk, not a quantified cost increase: the retained sources do not provide an average reasoning-token count or latency measurement for a particular workload.
The same migration analysis identifies the places a legacy string can survive beyond the main request code: environment variables, gateway rules, framework configuration, evaluation harnesses, scheduled scripts, queued payloads and failover paths. A separate report, published before the cutoff, likewise described the alias deadline as an immediate production concern and noted that deepseek-reasoner maps to Flash thinking mode rather than Pro.
For teams using the Anthropic-compatible interface, DeepSeek says the endpoint is supported, but the retained primary V4 notice does not provide enough implementation detail to prescribe an SDK-specific migration. The safe claim from the primary source is narrower: keep the relevant base URL and select a V4 model deliberately; verify the live provider guidance before changing production settings.
The unanswered question is not whether a replacement name exists. DeepSeek has supplied the compatibility mapping. It is whether an application’s selected tier and mode still meet its quality, latency, throughput and cost requirements after the alias disappears.
The next evidence should be a representative evaluation that records the chosen model and thinking setting, input cache status, input and output tokens, latency, error rate and concurrency behavior. That will establish whether Flash remains suitable and whether Pro’s higher cache-miss and output rates—and lower listed concurrency limit—buy a material improvement for the particular workload. Until then, treating Pro as an automatic “reasoner” upgrade would go beyond what the documented alias mapping supports.
Alphabet’s proposed $80 billion equity package is partly tied to AI infrastructure, but its $40 billion at-the-market program is primarily intended to cover employee-equity tax obligations. Together with Microsoft’s $190 billion 2026 capital-expenditure forecast, the disclosures show a more complicated shift in how hyperscalers fund constrained compute capacity.
Intel is putting €5 billion into equipment and connections at its operating Leixlip campus to increase output of Intel 3-based Xeon processors. The investment strengthens an existing European manufacturing base, but Intel has not disclosed its added wafer capacity, the split between its own products and foundry work, or an outside customer.
Sundar Pichai said Google supports an industry letter on open models, aligning the company with developers and technology firms resisting broad limits on Chinese AI systems while leaving allegations of IP theft and national-security risks unresolved.
Brazilian courts and bar associations have responded to hidden prompts in legal filings with fines, referrals, suspensions and an investigation. The known cases were detected, however, and do not show that an AI system changed a judicial decision; they instead test how courts secure AI-assisted review of untrusted case documents.
Salamanca City Central School District has put its proposed Realbotix humanoid robot and Optio AI assistant pilot on hold while it works on student-data privacy agreements. The pause leaves the district with a more basic task: establish what the digital tutor and the physical robot each collect, do and improve before either reaches students.
Samsung Electronics and Broadcom have signed an MOU covering memory, foundry technologies and advanced packaging for next-generation AI infrastructure. The companies expect more than $200 billion of collaboration over five years, but have not disclosed how that figure translates into purchases, production or capacity commitments.
OpenAI says GPT-5.6 Sol reached a 24.9% ExploitGym pass rate under a two-hour cap, up from GPT-5.5’s 15.1%. The benchmark measures a bounded step from a known vulnerability to an exploit; a separate assessment performed with OpenAI finds continued limits on hardened targets and long operations.
Chey Tae-won, chairman of SK Group, said Anthropic had sought SK Hynix component-supply cooperation for chips it hopes to make itself. But Anthropic has disclosed no chip plan, and the same summit also produced a separate letter of intent between Anthropic and SK Telecom on a large AI data center.
South Korea announced $950 billion in planned five-year semiconductor cooperation involving Samsung, SK Group and global technology companies, alongside prospective AI-data-center projects. The figure aggregates arrangements with different terms and disclosed commitments, leaving delivery, power and the overlap between the SK and NVIDIA figures unresolved.
OpenAI said it applied a fix after a July 25 disruption affected ChatGPT, its Codex coding service and parts of its API. The available reporting points to recovery in about 50 minutes, but the retained incident records do not explain the cause, the mitigation or whether the services failed through a common component.
Anthropic CEO Dario Amodei said the Claude developer has signed supply deals with Samsung Electronics and SK hynix. The available disclosures show a useful infrastructure relationship, but not an HBM purchase, a dedicated allocation or a claim on either supplier’s output.
Moody's Ratings expects six US hyperscalers to spend $785 billion in 2026 and close to $1 trillion in 2027. The forecast signals continued demand for AI capacity, but a separate Moody's analysis shows why capex alone understates the financing question: five of the companies already reported $969 billion in undiscounted future lease payments, much of it not yet on their balance sheets.
Anthropic has released Claude Opus 5 at Opus 4.8's API price and says it approaches Claude Fable 5 on coding and knowledge work. An Epoch AI listing places it at 159 ECI, below Fable 5 and GPT-5.6 Sol, leaving customers to test whether the cheaper model delivers better economics in their own workloads.
Nvidia plans to invest $1 billion in Naver as Brookfield signs a nonbinding term sheet for up to $9 billion toward a South Korean AI-factory project. The companies target a 200-megawatt expansion by 2028, but financing, definitive terms and construction remain ahead.
Nvidia and SK Group signed letters of intent for an AI-infrastructure initiative they value at more than $500 billion, centered on an up-to-2-gigawatt Korean AI factory and a long-term SK hynix memory partnership. The disclosed projects give the plan a technical shape, but not a capital breakdown, time horizon or basis for treating the headline figure as committed spending.
A transmission-line outage in Northern Virginia prompted data centers to transfer to backup power, removing about 3.1 GW of load in roughly 30 seconds. PJM reported no reliability impact, but the event put the operational behavior of concentrated computing campuses—not just their growing electricity demand—at the center of the grid-planning debate.
Imperial Valley Computer Manufacturing is asking a California court to overturn Imperial Irrigation District’s denial of water for a proposed 330-megawatt data center. The developer says it can offset cooling demand by fallowing nearby farmland, but IID policy places conserved water under the district’s control and requires a separate review of industrial service.
DeepSeek has reportedly told prospective investors that it will not sign second-round agreements in the coming days. The account describes a potentially temporary pause, leaving the size, valuation and timing of the proposed financing unconfirmed.
U.S. technology employers announced 139,156 job cuts through June 2026, 83% more than a year earlier. The figures coincide with huge planned data-center outlays, but they do not show how many jobs AI directly replaced—or whether the announced reductions become a lasting employment decline.
Emails show Virginia environmental officials began developing a critique of a health-impact study of Vantage Data Centers VA2 before it was published. The disagreement is less a single factual contradiction than a clash between a full-permit screening scenario and DEQ's estimate of the facility's reported 2025 operations.