Google's workhorse tier is shipping every three weeks while its frontier tier is two months past its announced date. The asymmetry is a predictable consequence of how the two are built, not a management failure — and for most production workloads it means the workhorse tier is now the product. Plus a scheduled 100% price increase with a published date.
On 13 August, Google released Gemini 3.7 Flash — three weeks after Gemini 3.6 Flash, and roughly nine weeks after the Flash release before that. On the same day, Gemini 3.5 Pro, the flagship Sundar Pichai told developers in May would arrive in June, was still not out. One company, two release clocks: the cheap tier turning every three weeks, the frontier tier stopped two months past its own announced date.
The pairing gets sharper when you read the stated reason for the delay. Bloomberg reported on 16 July, citing ten current and former employees, that Gemini 3.5 Pro is behind schedule because its coding performance falls short of internal expectations relative to OpenAI and Anthropic — and that Google reset and updated the model's underlying training data in late June to address it, with internal results that were again disappointing. Alphabet shares fell on the report. Coding is exactly the capability Google is now marketing the new Flash model on.
The easy read is that Google is in trouble. The more defensible read, and the one worth a buyer's time, is that the workhorse tier and the frontier tier have decoupled for structural reasons that are not going away — and that for most production workloads the workhorse tier is now the product.
Google's published evaluations put Gemini 3.7 Flash at 1588 Elo on Code Arena's web development leaderboard, against 1538 for its own 3.6 Flash, 1541 for Claude Sonnet 5 and 1523 for GPT-5.6 Terra. On DeepSWE v1.1 it reports a jump from 49.0% to 65.3%; on FrontierCode 1.1 Main, from 34.4% to 43.6%. Google describes it as its most intelligent workhorse model, aimed at coding, web development and agent workflows.
Every one of those numbers is Google evaluating Google's own model. None has been independently replicated. A 47-point Elo spread over Claude Sonnet 5 sits inside the range where harness choices, prompt scaffolding and sampling settings can move a result, and vendor leaderboards are not a neutral instrument. Read them as a claim about direction, not a settled ranking.
Even heavily discounted, the direction is informative. The gap Google claims between its own 3.6 and 3.7 Flash — 50 Elo, 16 points of DeepSWE — is larger than the gap it claims over two rival models from labs whose frontier tiers are shipping. A tier positioned eighteen months ago as the cheap fallback is being positioned now as the default.
A frontier model begins with a pre-training run: a single, months-long, capital-committed pass over an enormous corpus on a reserved cluster. You cannot partially roll it back. If the resulting base is weaker than you hoped on a capability you care about, your options are to accept it, to patch it in post-training, or to start again — and the third option costs you the calendar. Bloomberg's account of a late-June training-data reset is a description of exactly that decision being made.
A workhorse model is downstream of that base. Distillation, instruction tuning, reinforcement learning against task-specific reward signals, tool-use scaffolding — all of it operates on a base that already exists, on far smaller compute budgets, with far shorter feedback loops. Three weeks is a plausible cycle time for that work. Three weeks is not a plausible cycle time for a pre-training run, and no amount of organisational discipline makes it one.
So the cadence gap is not a management failure and it is not evidence that Google's frontier research is broken. It is the expected shape of a two-tier release strategy while the base underneath is being reworked. Every lab with the same structure has the same asymmetry. Google is simply the one where both clocks happen to be visible at once.
The corollary is less comfortable. If the workhorse tier keeps improving on a three-week cycle while the flagship stays in the shop, the flagship's eventual advantage has to be large enough to justify both the wait and the price gap. Each Flash release raises that bar from underneath.
Hold two facts together without over-reading them. Google is selling 3.7 Flash primarily on coding and agentic workflows. Google's flagship is reportedly delayed primarily because of coding. It is tempting to conclude that the first explains the second — that resources went to the cheap tier at the frontier tier's expense. Nothing in the public record supports that.
A more mundane explanation fits better. A distilled model can be very good at the coding tasks benchmarks measure — well-specified problems with checkable outputs and bounded context — while a frontier model is being held for capabilities that show up on longer, messier, less checkable work. The evaluations Flash is winning are not obviously the evaluations holding Pro. Both statements about coding can be true and describe different things.
Reporting has also cited senior researcher departures and the possibility of a retrain from pre-training. Google has not confirmed either, and nothing above depends on them. Note them; do not build on them.
Introductory pricing is $0.75 per million input tokens and $3.75 per million output, through 31 December 2026. On 1 January 2027 it becomes $1.50 and $7.50 — exactly double, on both sides, on a date Google has already published.
This is not the hypothetical repricing risk we wrote about when DeepSeek raised API prices by up to 1,100% earlier this month. It is a scheduled one, disclosed in advance, with a date attached. That makes it easier to plan around and correspondingly easier to ignore. A team that models unit economics on the launch rate and ships in Q4 has a 100% input-cost increase arriving in the middle of its first full quarter of production traffic.
The practical move here is arithmetic, not strategy: run the cost model at $1.50 and $7.50, and treat the 2026 discount as a margin windfall rather than the baseline. If the workload only clears at introductory pricing, it does not clear.
The framing that survives all of this is the one about tiers. For several years the sensible default was to reach for the frontier model and fall back to the cheap one when cost bit. On the current evidence — vendor-reported, unreplicated, but pointing consistently in one direction — that ordering has inverted for a large class of production work. The frontier tier is where you go when the workhorse demonstrably fails, and it is no longer clear how often that happens.
Google's own evaluations say yes by a narrow margin — 1588 Elo on Code Arena web development against 1541 for Claude Sonnet 5 and 1523 for GPT-5.6 Terra. Those are vendor-run numbers with no independent replication, and the margin is small enough that evaluation harness and prompting choices could account for it. On your specific workload, the honest answer is that you have to test it yourself.
DrafterDaily covers AI, business and markets with the second-order consequences included. One analytical brief a day, no hype.
Browse AI coverageClaude Fable 5.1 costs exactly what Fable 5 cost per token. The 25-to-45% saving Anthropic advertises comes from one repriced line item — cached input, now billed at 2.5% of list instead of 10%. That is a discount you only collect if you keep the agent running.
Anthropic was not breached. The malware was already on the customer's machine, and it took a session cookie rather than a password — which is why two-factor authentication did nothing and why server-side revocation is the only lever the vendor has.
OpenAI's incident report on the Hugging Face breach leads on a security failure. The remarkable part is a verbatim chain-of-thought in which a model identifies its action as unauthorised and proceeds anyway — and an escape route that was a package manager, not a superintelligence.