Google's New Pro Model Still on Hold, Discount on Gemini 3.7 Flash to Bridge the Gap

Google distracts from missing Pro model with a limited discount on Gemini 3.7 Flash

On Thursday, Google released Gemini 3.7 Flash, a model focused on coding and agentic workflows, replacing version 3.6 Flash which was only three weeks old. The company did not provide a release date for its flagship model, Gemini 3.5 Pro, which has been announced for months, as reported by Reuters and Axios. Google had stated in July that 3.5 Pro was being tested with partners and would be coming “soon”; at Google I/O in May, June was still mentioned as the launch date, according to Ars Technica. Google attributes the rapid release cycle to developer feedback and algorithmic improvements.

Until the end of 2026, Gemini 3.7 Flash will cost $0.75 per million input tokens and $3.75 per million output tokens, half the original introductory price of 3.6 Flash. Starting January 1, 2027, the rates will increase to $1.50 and $7.50. According to The New Stack, Google also applied this discount to 3.6 Flash until December 31.

Technically, the model offers three thinking_level tiers (low, medium, high), with medium being the default; lower tiers target latency-critical applications, while high is for difficult coding, math, and planning tasks. The token consumption over the entire agent loop thus determines the actual Inference costs. 3.7 Flash is available in AI Studio, Android Studio, Google Antigravity, the Gemini Enterprise Agent Platform, and for end-users via the Spark agent for AI Pro and Ultra subscribers in more than 160 countries. According to the provider, the protective measures against misuse in the areas of CBRN and cyber-attacks have also been updated.

The launch comes during a turbulent period at Google DeepMind. Last week, Demis Hassabis handed over leadership to his deputy, Koray Kavukcuoglu, while at the same time, the two original technical co-leads of Gemini left the company to found their own startup. Reuters also reported that co-founder Sergey Brin urged key AI employees to bet everything on Gemini. Sundar Pichai had rejected criticism of the flagship model’s delay during the July earnings call. → VentureBeat, The New Stack, Google, Thurrott, Quartz, Finimize, Bloomberg, Business Standard, 9to5Google, Ars Technica

Synthszr Take: 65.3 percent on DeepSWE, up from 49, is the number currently circulating through developer channels and deciding model choices. The second number is in the fine print and isn’t shared in any benchmark screenshot: On January 1, 2027, Google will double the rates to $1.50 and $7.50 per million tokens, and this will hit the teams hardest that are now setting the model to thinking_level high to really solve the tricky cases. An agent that burns through ten tool calls and a lot of context per task looks like a bargain in August and like an unplanned budget item in January. The sensible reaction is unspectacular: calculate the cost per completed task using the 2027 rate, not the introductory price, and keep the model call behind a configuration that can be switched without refactoring.

OpenAI releases Ultrafast mode for GPT-5.6 Sol with 750 tokens per second

OpenAI has released a new operating mode called Ultrafast, which, according to the company, allows the current frontier model GPT-5.6 Sol to run at 14 times the speed of standard processing. Specifically, OpenAI mentions up to 750 output Token per second. In a blog post on Thursday, the company wrote that real-time speed has so far usually meant switching to a smaller or specialized model. Technically, the mode runs via OpenAI’s partnership with chip manufacturer Cerebras. Ultrafast is currently available as a preview and is only enabled for a small group of customers. OpenAI announced it will expand access as capacity grows. A price for the mode is not mentioned in the announcement. → TechCrunch

Synthszr Take: 750 tokens per second sounds like a feature, but it’s a budget issue. Speed is paid for in compute, and Cerebras’s capacity is obviously limited, otherwise OpenAI wouldn’t have to cap the preview to a small group of customers. The one number missing from the announcement is the price per token in this mode, and without it, any calculation is just guesswork. A hard token limit per run and a metric of “cost per resolved ticket” instead of “tokens per second” should be part of the setup before the first Ultrafast job starts. For incident response, where a minute saved is real money, the premium probably pays off immediately; for customer service-Batches the question only becomes interesting once OpenAI provides the price list.

DeepSeek releases V4 Pro and keeps the price at 43.5 cents per million tokens

DeepSeek has moved its flagship model V4 Pro from the preview phase to general availability without raising prices. A call still costs $0.435 per million input tokens and $0.87 per million output tokens, a fraction of what Fable 5 costs, according to The Code. The model works with a context window of one million tokens and supports Tool Calling, allowing it to be directly integrated into existing agent pipelines. In agentic coding, V4 Pro is still slightly behind Fable 5, according to the provider. → The Code

Synthszr Take: The real story here is that the price didn’t increase when moving from Preview to General Availability. The opposite would be typical: Preview is cheap, GA is more expensive once customers have built their pipelines. DeepSeek is forgoing this surcharge, thereby marking 43.5 cents as the floor price for near-frontier coding. For any smaller provider whose business model is built on a few dollars of markup per million tokens, this is an uncomfortable calculation, because at a quarter of the Grok rate and a fraction of the Fable 5 rate, there’s no room left for “we are the cheap alternative.” What remains are reasons that have nothing to do with the token price: European Hosting, audited domain data, latency guarantees, liability.

US intelligence agency calls for digital birth certificates for AI agents

Doug Cossa, Chief Information Officer of the U.S. Intelligence Community at the Office of the Director of National Intelligence, called for a unified identity system for AI agents at the Defense Intelligence Agency’s DoDIIS conference. According to his appearance, reported by Breaking Defense and covered by PYMNTS, agentic artificial intelligence turns the principle of Zero Trust on its head: instead of minimal rights assignment, an agent gets everything it needs to work independently. Cossa argues that agents, like users and devices, request, store, and process data and therefore require the same identity foundation. This identity is intended to form the basis for deciding what an agent is allowed to do at all. According to him, a cross-agency system for establishing and controlling the identity of autonomous bots does not yet exist.

Synthszr Take: Liability needs an addressee, and currently, almost no agent architecture provides one. In practice, an agent runs under an employee’s access or a shared service account, and when things go wrong, the log shows an identifier that could point to two teams, a provider, and a forgotten script. Cossa’s call for a common identity system is essentially a question of evidence: without identity, no attribution; without attribution, no accountability. The reassuring thought that a human is always liable in the end only holds up until one agent launches a second agent and passes on its rights without any documentation of on whose behalf it happened. What matters now is a dedicated entitlement principal per agent, with an expiration date, a named client, and a log that makes the chain readable backwards.

Addy Osmani: Automated Guardrails Replace Code Review for AI Agents

Addy Osmani argues in his Substack newsletter that classic code review is reaching its limits with AI agents because simply too much code is being generated for anyone to read. Quality assurance is therefore shifting to the agent harness, the environment, and the operating system surrounding the agent. Constraints define what a system is allowed to do: tests and deterministic checks run against every suggestion from the agent before it is approved. Osmani describes scenarios in which agents generate hundreds of thousands to millions of changes per day, insisting that only such loops can reliably deliver production-ready software. → Substack

Synthszr Take: Two reviews, three answers, the same piece of code. This is the real message in this debate, and it dismisses the idea that the human reviewer can simply be replaced by a second agent. A merge cannot depend on a judgment that might be different on the next run, so the review instance must be deterministic: tests, types, contract checks, reproducible builds, hard gates on every commit. This shifts the most expensive work in the development team from reading to building the verification mechanics, and for ten years, precisely this work was considered a cost center to be cut in the next round of savings. When agents produce hundreds of thousands of changes a day, the testing infrastructure is what makes speed usable in the first place; without it, any acceleration remains a hope that has to be cleaned up in production.

DeepSeek Harness (dsh) is Now Open Source

On Thursday, DeepSeek open-sourced the DeepSeek Harness (dsh), a Node.js-based runtime environment for coding agents under the MIT license. According to The New Stack, the GitHub repository gathered over 33,000 stars within a few hours, and a catalog of community plugins is already forming. The project is explicitly labeled as a Developer Preview; in the README, DeepSeek announces breaking changes. The web interface is started via npx @deepseek-ai/dsh web on port 3080. Architecturally, the agent harness follows the principle that everything is a plugin, and the team means this literally: model adapters, the tool registry, the session log, and the agent loop itself are all exchangeable plugins. The documentation describes a system with no privileged core that would need to be patched. The foundation is Cordis, a meta-framework for dynamic composition, whose design is described in a paper by three researchers from Peking University and DeepSeek. → The New Stack, GitHub

Synthszr Take: What the model sees must be reconstructible from a single event stream. This makes traceability a design requirement, with resume, fork, and replay becoming byproducts. The downside is also evident in the repo: 33,000 stars in a few hours, a rapidly growing third-party plugin catalog, and a preview status that promises breaking changes. Before any production use, it must be clarified which plugins are in the loop, who maintains them, and whether the event stream is complete. Landlock and Seatbelt protect the file system, but they don’t protect the plugin supply chain.

Apple Wants to Pay Publishers Per Siri Request Instead of a Flat-Rate License

According to a report by The Wall Street Journal, Apple is negotiating with publishers to pay them for the use of current news content in the revamped Siri. Instead of the industry-standard flat-rate license, which covers broad access to an archive, Apple has proposed a usage-based model: payment will be made each time Siri uses material from a publisher. For these payments, Apple is reportedly considering a budget in the nine-figure range. → Techpresso

Synthszr Take: With the per-retrieval model, timeliness gets a unit price for the first time, and it will apply across the entire market. Previously, publishers received flat fees that looked good in their annual planning but obscured the actual usage value. A nine-figure budget sounds like a lot, but it’s distributed over every single request, and after six months of operation, Apple will know exactly which source is used how often and which is never used. This means that in the future, every assistant provider will buy freshness like it buys computing time: variable, measurable, and negotiable downwards. For publishers, this means that compensation will be linked to demand rather than brand name, and most will not win in this scenario.

German Researchers Find API Keys in Agent Logs

A research team from Germany has decoded 315,320 encrypted Reasoning Blocks from 6,708 publicly shared agent trajectories and found private data that did not appear in the visible chat histories. The trajectories came from GitHub and Hugging Face, and the commercial interfaces from Anthropic, OpenAI, and Google were affected. The basis is a preprint from August 10, 2026, co-authored by, among others, Alexander Panfilov from the ELLIS Institute Tübingen and the Max Planck Institute for Intelligent Systems. The method did not crack any encryption: an intact block was presented to a compatible, less-protected model from the same provider, which accepted it as previous thought steps and transcribed it upon request. Preliminary work on this was done by cryptographer Matthew Green from Johns Hopkins University, who showed in May 2026 that such blocks could be reused across sessions and accounts. → Marcus Schuler

Synthszr Take: In every developer team, agent logs circulate like screenshots: quickly pasted into an issue, committed to the repo, uploaded to Hugging Face as a dataset so someone can reproduce the error. This is exactly where the researchers collected 6,708 trajectories, and in 328 sessions, or 4.9 percent of the sample, at least one real secret was found. The really uncomfortable finding is the 64 artifacts that were only in the hidden reasoning: when cleaning up the conversation, the model removed the passwords from the readable text but left them in the encrypted attachment that no one can check. This renders any visual inspection before sharing useless, and the usual secret scanners in the commit hook also fail because they see the block as binary noise. The patch from the providers eliminates one attack vector, but the logs remain openly available on the internet, and keys do not expire over time.

Near-Autonomous AI Attack on Taiwan’s Government: 2,500 Personal Data Records Leaked

Suspected Chinese attackers have compromised a Taiwanese government agency using freely available AI models, exfiltrating over 2,500 personal data records. This was reported by the Israeli security firm Dream in a blog post published on Wednesday, which was first covered by the Financial Times. According to Dream, it is a “multi-agent system”, based on the open-source frameworks Hermes and OpenClaw, and was configured, according to the researchers, to adapt mid-operation without human intervention. The firm describes so-called “Learning Cycles” in which the system autonomously searches vulnerability databases, GitHub repositories, and security publications for suitable techniques for the target’s infrastructure. The operation did not stop at the original target: according to Dream, government IT service providers, a nuclear safety authority, a government email system, and more than seven energy companies were simultaneously scanned for misconfigurations and open admin interfaces. The safeguards of the models were bypassed by the attackers, according to Dream, by declaring their requests as an authorized penetration test. → Techpresso

Synthszr Take: The difference between automated and autonomous can be pinned down to one point: the question of who decides the next step. A scanning script works through a list that someone wrote beforehand. The system Dream uncovered prioritizes its targets Bayesianly, self-corrects after failed attempts, and enters its own research cycles when it lacks the necessary techniques. This is exactly what Dream’s own assessment states, and it’s the most interesting part of the entire report: simply running a model isn’t enough; it requires coordinated agent orchestration and fine-tuned decision logic. The 1,400 files in the open archive are the real piece of work; the model is the cheapest part of it.