WSJ reports Google is preparing a Flash-tier model — not a frontier one — to close its coding gap with Anthropic, and the choice of tier is the whole story.
Flash Cadence — WSJ-Reported, Not Official
Jul 21, 2026
Gemini 3.6 Flash ships — first data point in the accelerating Flash cadence
Reporting by Erin Woo at The Wall Street Journal — picked up by Investing.com, Yahoo Finance, and Gizmodo — says Google is preparing a model called Gemini 3.8 Flash, internally code-named Skimaki, with a focus on coding performance and a potential release around September 2. The report cites people familiar with the matter; Google has not confirmed the model, issued no blog post, and as of this writing the Gemini API documentation still lists 3.7 Flash as the latest Flash model with no 3.8 Flash model ID. Readers should check whether Skimaki has actually shipped before treating any of this as settled fact — a reported “as soon as Wednesday” either lands within hours or ages quickly.
According to the WSJ report, Google employees testing Skimaki inside an internal coding platform code-named Jetski preferred it to Anthropic’s Opus on coding tasks in side-by-side comparisons. Three guardrails belong on that sentence before it travels further. First, it is a human-preference signal from Google’s own staff on Google’s own tool, reported through unnamed sources — not a third-party benchmark; no SWE-bench or LiveCodeBench score exists for this model, and any figure circulating comes from leak communities, not the WSJ report. Second, “preferred to Opus” is a preference test, not a capability-equivalence claim; in Flash-versus-frontier comparisons, latency and output concision can drive preference as much as raw reasoning. Third, Skimaki and Opus do not occupy the same pricing tier — this is not evidence that Google has matched Anthropic at the frontier. Skimaki is the model; Jetski is the internal tool — the two names are already being swapped in secondary coverage.
Separately and on the record, Google CEO Sundar Pichai has acknowledged that on agentic coding — tool use, instruction following, long-horizon tasks — Google is “a bit behind at this moment,” attributing part of the gap to Google’s lack of a developer-facing coding surface generating a data flywheel, the exact advantage Anthropic has built by distributing through third-party developer surfaces like Cursor.
The key insight: Google’s reported answer to Anthropic’s coding lead is not a bigger frontier model. It is a Flash-tier model — small, fast, cheap, and fast to modify. That choice of tier is the entire strategic argument, and it is a sharper one than the headline implies.
The instinct when you are behind on a capability is to build a bigger model. The reported logic behind Skimaki runs the opposite direction, and the rationale given in the WSJ piece is worth taking seriously on its own terms: a smaller model requires far less compute to modify, which means multiple research teams can run reinforcement-learning iterations on it in parallel. The organization gets to try more things per week. Google is not, if this framing is right, buying capability through parameter count — it is buying iteration velocity and unit economics.
The cadence makes that argument concrete. Three Flash releases in roughly six weeks, on approximately a three-week rhythm — 3.6 on July 21, 3.7 on August 13, now 3.8 reportedly imminent — against a stated goal of roughly monthly model updates. That is a release clock, not a moonshot schedule. It matches a pattern visible across the rest of this week’s competitive dynamics: the frontier is moving to cost, and the interesting engineering is making a good-enough model cheap to run and fast to improve, not making the single smartest model.
Sundar Pichai — Hard Fork / SEJ (On the Record)
“A bit behind at this moment” on agentic coding — and the diagnosis he offers is not model quality but data: Google lacked the developer-facing coding surface that generates the flywheel Anthropic acquired by distributing through surfaces like Cursor.
That data-flywheel diagnosis is the other half of the structural picture. Anthropic’s coding lead is not purely a model-quality story — it is partly a distribution story. Distributing through third-party developer tools like Cursor put Anthropic’s models in front of working developers doing real tasks, generating the feedback signal that sharpens coding performance. Google, building its own internal surfaces, did not have that loop. Skimaki — if the report is right — is the counter-move: not a bigger brain aimed at the frontier, but a cheaper, faster, more-iterable one aimed squarely at the margin of the market where coding work actually gets done.
Margin-Structure Risk: The Premium Gets Repriced From Below
Anthropic’s coding lead is today priced at frontier rates — enterprises pay premium per-token prices for Claude because it is the best at code. If a Flash-tier model genuinely rivals Opus on real coding work — and that is a large conditional resting on an internal preference test — the price of that capability collapses. A cheap model does not just compete for the win; it removes the premium. That is a different and more structurally dangerous threat than a better frontier model. This claim remains explicitly conditional: the margin only breaks if Skimaki holds up on real coding tasks, which an internal preference test cannot establish.
If the reported rationale holds, Google has decided that the way to close a capability gap is not to throw more compute at a bigger model but to make a smaller model cheap enough to iterate on at speed. Three Flash releases in roughly six weeks is not a product launch rhythm — it is a compounding-improvement rhythm. The organization that can run more RL experiments per week wins on the time axis, not the scale axis. Whether that bet pays off on coding specifically is unproven, but the logic is transferable to every capability gap Google faces.
IMPLICATION 2 — Anthropic’s Coding Premium Is the Asset Most Exposed to Cost Compression
Anthropic’s strongest enterprise position right now is coding, and coding is priced at frontier rates. A Flash-tier model that genuinely competes on coding tasks — conditionally — does not just steal share; it reprices the category. Enterprises evaluating Claude for code will now have a credible cost-structure question to ask. Anthropic’s defensible response is the data flywheel: more real developer usage through Cursor and similar surfaces generating better signal faster. That flywheel advantage does not evaporate overnight, but it needs to compound faster than Google’s iteration cadence.
IMPLICATION 3 — The Flash Tier Is Now a Strategic Axis, Not Just a Price Point
Flash models were initially positioned as cheaper, faster versions of frontier models — good enough for high-volume, latency-sensitive tasks, not for hard reasoning. The Skimaki framing, if accurate, repositions Flash as the place where Google is actively trying to win on capability, not just cost. That is a meaningful shift in how to read the competitive map: the interesting capability battle may no longer be at the top of the model stack. It may be in the middle tier, where unit economics let one company iterate faster than another can build bigger.
Business Engineer Framework
The Map of AI Redrawn — Where Skimaki Fits in the Stack
The Map of AI tracks 200+ companies across nine layers of the AI stack. The Skimaki story sits at the intersection of two layers: the model tier (Flash vs. frontier) and the distribution layer (who has the developer surface generating the data flywheel). Understanding which layer is actually under competitive pressure — and from which direction — is the read that separates a headline from a strategic signal. The full map is the tool for that read.
The WSJ report — unconfirmed by Google, describing an unshipped model, citing an internal preference test that is not a benchmark — needs every one of its hedges kept intact. What survives the hedges is the shape of the bet: Google’s reported response to being behind on coding is not a bigger model but a cheaper, faster, more-iterable one, aimed at the margin rather than the frontier. If Skimaki holds up on real coding tasks, the most valuable-looking lead in enterprise AI today — Anthropic’s coding premium — turns out to be the one most exposed to being repriced from below. Whether that model is Skimaki is unproven. That Google is aiming at the margin rather than the frontier is the story worth marking.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
