A Wall Street Journal report that Google might ship a new Gemini model “as soon as Wednesday” did something unusual for an AI rumor: it moved a trillion-dollar stock. Alphabet shares climbed roughly 0.7% in after-hours trading once the report crossed the wire, clawing back part of a nearly 1.3% slide during the regular session. The model in question, Gemini 3.8 Flash, carried the internal codename “Skimaki” and was reportedly aimed squarely at closing Google’s coding gap with OpenAI and Anthropic.
That report, covered by Yahoo Finance under the headline “Google Could Release New 3.8 Flash AI Model As Soon As Wednesday, Says Report — GOOGL Stock Gains After-Hours,” has kept circulating through trading desks and AI forums for weeks. Gemini 3.8 Flash has since actually shipped, which means the market now has something it rarely gets with AI rumors: a chance to check the hype against the scoreboard. The early verdict is mixed, and that nuance is what makes this story worth revisiting rather than filing away.
This piece walks through what the report claimed, what Google’s engineers reportedly found when they tested the model internally, what independent benchmarks now show, and why the market’s muted reaction may say more about AI fatigue among investors than about the model itself.
What the Wall Street Journal Report Actually Said
According to people familiar with the matter cited in the original report, Google was weighing a release of Gemini 3.8 Flash within days of the story breaking. The report did not confirm a release date with certainty. It described Wednesday as a possibility, not a lock, and left plenty of room for Google to slip the timeline, which is standard practice for a company that has been burned before by over-promising on AI ship dates.
The more interesting detail buried in the reporting was not the date. It was the claim that Google’s own engineers, testing the new model through an internal coding tool known as Jetski, preferred Gemini 3.8 Flash’s output to Anthropic’s Opus model in head-to-head coding comparisons. That claim came from sources familiar with the testing rather than an on-record Google statement, so it should be read as an internal signal rather than an independently verified benchmark. Still, for a company that has spent much of 2026 fending off the perception that it trails OpenAI and Anthropic on code generation, that kind of internal preference, even unverified, was newsworthy enough to move a stock with a market capitalization in the trillions.
Why a 0.7% After-Hours Bump Is a Bigger Deal Than It Looks
A 0.7% after-hours move sounds small next to the swings AI stocks have pulled off this year. But context matters. Alphabet closed the regular session nearly 1.3% lower before the report landed, meaning the after-hours gain represented a genuine reversal in sentiment, not just noise on top of an already-green day. For a company the size of Alphabet, even fractional percentage moves translate into tens of billions of dollars in market value shifting in minutes, all triggered by a report that explicitly hedged on whether the thing it was describing would even happen that week.
That is the pattern that has defined AI-stock trading through 2026: headlines about unreleased models move prices faster than the models themselves ever could. Traders are no longer waiting for a launch. They are pricing in the probability of a launch based on leaked internal codenames and secondhand descriptions of internal testing tools. Gemini 3.8 Flash’s after-hours bump is a clean case study in that dynamic, and it is one reason the story kept getting cited long after the model actually shipped.
Inside “Skimaki”: What We Know About the Codename
Google reportedly referred to the project internally as Skimaki before it carried the public Gemini 3.8 Flash branding. Internal codenames at Google’s DeepMind division rarely survive contact with a public launch, and this one did not either. What is notable is how much the codename leaked before the product did, which has become something of a pattern this year across the major labs. Model names now seem to leak through engineering forums, internal testing logs, and chatter among contractors well before the companies themselves are ready to confirm anything on the record.
Google has not published an official account of why “Skimaki” was chosen or dropped, and no source in the original reporting offers a definitive explanation. That is worth flagging plainly: anything beyond the codename’s existence and its eventual replacement by the Gemini 3.8 Flash name moves from reported fact into speculation, and this piece will not go there.
Jetski: The Internal Tool Behind the Opus Comparison
Jetski is described in the reporting as Google’s internal coding evaluation tool, the harness engineers reportedly used to run head-to-head comparisons between candidate models before deciding what ships. The claim that Gemini 3.8 Flash came out ahead of Anthropic’s Opus model in those internal tests is attributed to people familiar with the testing, not to a published Google benchmark. That distinction matters for anyone trying to size up the real competitive gap between Google and Anthropic right now, a question this site has tracked through prediction-market data that shows the two labs running close to even in trader sentiment.
Internal preference tests are useful signals, but they are not blind, third-party benchmarks. Companies test against the versions of rival models they have access to at the time, under conditions they control, and they have every incentive to publicize results that favor them. None of that means the Jetski result was wrong. It just means outside observers should treat it as directional evidence rather than as proof that Gemini 3.8 Flash beats Opus across the board, especially once Anthropic’s own newer releases enter the picture.
The Benchmarks That Came After the Headlines
Once Gemini 3.8 Flash actually shipped, independent evaluators got a chance to run their own numbers, and those numbers give a clearer picture than the pre-launch rumor ever could. On the Artificial Analysis Intelligence Index, a composite benchmark widely cited across the AI industry, Gemini 3.8 Flash scored 41 at high reasoning effort, according to Artificial Analysis, which does not publish a comparable score for Gemini 3.7 Flash. That score lands in the same range as GPT-5.6 Sol at extra-high reasoning effort, according to comparison data published by Kingy AI’s 192-run work-per-dollar test, which pitted the two models against each other directly across a large batch of tasks.
On HLE-Verified, the verified subset of Humanity’s Last Exam that strips out ambiguous or disputed questions, Gemini 3.8 Flash scored 54.9%, placing it inside the top 10% of models tracked on the HLE-Verified leaderboard as of October 2026, ranked 1st out of the 6 models tracked. Output speed came in around 302 tokens per second in that same leaderboard compilation, respectable for a model marketed as a fast, cost-efficient tier rather than a frontier reasoning flagship.
Where Gemini 3.8 Flash Wins and Where It Still Trails
Google’s own benchmark disclosures, cited in benchmark-aggregation coverage of the launch, painted a split picture rather than a clean sweep. Gemini 3.8 Flash reportedly led both Claude Opus and GPT-5.6 Sol on tests like Vals Finance Agent v2 and Terminal-Bench 2.1, two evaluations that lean on structured, tool-using tasks. On tougher, more open-ended benchmarks like Terminal-Bench 4.0 and OSWorld-2.0, which push models through longer agentic workflows with more room for drift and error, Gemini 3.8 Flash reportedly trailed its rivals.
| Benchmark | Gemini 3.8 Flash Result | Category |
|---|---|---|
| Artificial Analysis Intelligence Index (high effort) | 41 (Artificial Analysis does not publish a comparable score for Gemini 3.7 Flash) | Composite reasoning |
| HLE-Verified | 54.9% (ranked 1st of 6 models tracked) | Expert-level knowledge |
| Output speed | ~302 tokens/second | Throughput |
| Vals Finance Agent v2 | Leads Claude Opus and GPT-5.6 Sol | Structured agentic tasks |
| Terminal-Bench 2.1 | Leads Claude Opus and GPT-5.6 Sol | Coding / tool use |
| Terminal-Bench 4.0 | Trails Claude Opus and GPT-5.6 Sol | Long-horizon agentic coding |
| OSWorld-2.0 | Trails Claude Opus and GPT-5.6 Sol | Open-ended computer use |
That split result is a more honest story than a simple Google-beats-Anthropic or Google-falls-short-again framing. The model is genuinely strong on structured, well-defined tasks and genuinely weaker on the kind of long, messy, multi-step agentic work that has become the next frontier for coding assistants. Buyers choosing between Gemini 3.8 Flash, GPT-5.6 Sol, and Claude Opus should weigh which category of task dominates their actual workload rather than leaning on a single headline score.
Pricing: The Sticker Stayed Flat, the Real Cost Did Not
Google kept Gemini 3.8 Flash’s published pricing identical to Gemini 3.7 Flash: $0.75 per million input tokens and $3.75 per million output tokens. On paper, that looks like a free upgrade. In practice, it is not quite that simple. Because the new model reasons more deeply by default, it tends to consume more compute per completed task, and independent cost analysis pegged the real cost per completed task at roughly 40% higher than its predecessor despite the identical per-token sticker price.
That is a pattern worth watching across the industry generally. Published per-token pricing has become a less reliable guide to what companies actually pay, because reasoning-heavy models burn through more tokens to reach the same answer. Teams evaluating a swap from Gemini 3.7 Flash to 3.8 Flash should budget for that gap rather than assuming flat pricing means flat spend, a lesson that echoes what this site covered when Google restructured its consumer Gemini tiers earlier this year around similar cost-to-serve pressure.
Historical Context: Google’s Flash Line Has Always Been the Pressure Valve
Gemini’s Flash tier has never been Google’s showcase model. That job belongs to the Pro and Ultra tiers, and increasingly to Gemini 4 Argon, which Google pushed into a global rollout with a sharply expanded context window. Flash exists to do something different: prove that Google can ship a fast, cheap, good-enough model at a pace that keeps developers from defecting to OpenAI’s smaller tiers or Anthropic’s Haiku line. Judged against that mission, Gemini 3.8 Flash’s incremental three-point bump on the Artificial Analysis index and its mixed benchmark split are less a disappointment and more business as usual for a mid-cycle refresh.
| Model | AA Intelligence Index | Input / Output Price per 1M Tokens | Status |
|---|---|---|---|
| Gemini 3.7 Flash | 56 | $0.75 / $3.75 | Superseded |
| Gemini 3.8 Flash | 41 | $0.75 / $3.75 | Current Flash tier |
| GPT-5.6 Sol (extra-high effort) | 47 (not matched; Gemini 3.8 Flash scores 41) | Not disclosed in cited reporting | Rival flagship-tier model |
| Claude Opus (per internal Jetski testing) | Not independently disclosed | Not disclosed in cited reporting | Rival, reportedly trailed in some internal coding tests |
The repeated pattern across Gemini generations is incremental, closely-priced updates that trade places with rivals benchmark by benchmark rather than any single model lapping the field. That is a very different narrative from the one that circulates in headlines, where every release gets framed as a decisive move in a two-horse or three-horse race. The data suggests something closer to a treadmill, where OpenAI, Google, and Anthropic trade small leads across different task categories every few weeks.
Competitive Pressure on Anthropic and OpenAI
Even an unverified internal preference for Gemini 3.8 Flash over Claude Opus carries weight because of where Anthropic sits competitively right now. This site has covered how Anthropic has been leaning on cost cuts to keep Claude competitive on price even as rumors build around a future Opus 6 release. A report, even a hedged one, that Google’s own engineers preferred a mid-tier Gemini model to Anthropic’s flagship in coding tasks adds pressure on Anthropic to show, not just claim, that Opus holds its coding edge.
OpenAI faces a parallel pressure point. GPT-5.6 Sol matching Gemini 3.8 Flash’s score on the Artificial Analysis Intelligence Index at extra-high reasoning effort is, in one reading, a wash. In another reading, it means a Google model priced and marketed as a fast, lower-cost tier is matching a model OpenAI runs at its highest and presumably most expensive reasoning setting. If that comparison holds up under more independent testing, it is arguably a bigger problem for OpenAI’s pricing position than for Google’s.
Market Impact Beyond the Immediate Stock Move
The 0.7% after-hours bump was modest, but modest is relative in a market where AI headlines have repeatedly triggered multi-percent swings in chipmakers, cloud providers, and the labs themselves. The muted reaction here may reflect something other than indifference to the news. It may reflect that investors have started pricing AI model releases the way they price routine product updates from mature software companies, rather than as speculative breakthrough events. That shift, if it holds, would mark a meaningful change in how AI-adjacent stocks trade going into 2027.
It is also worth noting what did not move: Alphabet’s broader AI infrastructure narrative, including its custom silicon push, stayed on its own track regardless of Flash-tier model news. Readers tracking that side of Google’s AI strategy can see how it diverges from pure model-release headlines in this site’s coverage of Google’s Gemini Agent rollout, which drew its own distinct market reaction tied to product integration rather than raw benchmark scores.
What Enterprise and Developer Teams Should Actually Watch
For teams building on Gemini’s API rather than trading Alphabet stock, the practical takeaways are narrower and more useful than the market noise. First, the per-task cost increase matters more than the flat per-token price; budget for it. Second, the benchmark split between structured and open-ended agentic tasks means teams should test Gemini 3.8 Flash against their own actual workloads rather than trusting a single aggregate score. Third, Google’s official Gemini API release notes remain the most reliable source for what actually shipped, as opposed to what leaked ahead of launch.
Teams already standardized on Gemini 3.7 Flash have less urgency to switch than the headlines imply. A three-point composite benchmark gain paired with a roughly 40% higher effective cost per task is not an obvious upgrade for every workload, particularly for high-volume, cost-sensitive applications where Flash’s whole appeal was being the cheap option.
Predictions: Where This Goes From Here
- Expect Google to keep the Flash tier on a faster refresh cadence than Pro or Ultra, using it as the line where it tests reasoning upgrades before committing them to flagship models.
- Expect more internal-testing leaks like the Jetski and Opus comparison, since labs have shown little ability to keep pre-launch evaluation chatter contained in 2026.
- Expect pricing pressure to intensify as more labs face the same gap between flat per-token prices and rising per-task compute costs, pushing eventual public price adjustments across at least one major provider within the next two quarters.
- Expect Anthropic to respond with its own benchmark disclosures aimed specifically at agentic, long-horizon coding tasks, the category where Gemini 3.8 Flash reportedly struggled most.
- Expect market reactions to AI model releases to keep shrinking in magnitude as investors treat them as routine rather than extraordinary, unless a release demonstrates a genuinely discontinuous capability jump rather than an incremental benchmark gain.
The Bigger Picture: A Rumor That Outlived Its News Cycle
What makes this story worth a second look weeks after the fact is not the stock move, which was small, or the launch itself, which followed a familiar pattern. It is the gap between how the story was covered in the moment, as a market-moving leak about an unreleased model, and how it reads now, as a documented, benchmarked release with a mixed but measurable track record. That gap is becoming the norm for AI coverage generally: the leak generates the headline, and the actual performance data arrives later, quieter, and far more useful for anyone trying to make a real decision about which model to build on. Readers who want the full model-lineage picture can check Presenc AI’s Gemini model lineage tracker, which maps how 3.8 Flash fits into Google’s broader release sequence through 2026.
For Alphabet investors, the lesson is that a 0.7% after-hours pop tied to an unconfirmed report is not a signal to chase. For developers, the lesson is that the benchmark table matters more than the leak that preceded it. And for the AI industry generally, Gemini 3.8 Flash is another data point in a year defined less by any single breakthrough model than by a steady grind of mid-sized upgrades, each one nudging the competitive balance a few points in one direction before the next release nudges it back.
Frequently Asked Questions
What is Gemini 3.8 Flash?
Gemini 3.8 Flash is Google’s updated fast-tier AI model, reported internally under the codename “Skimaki” ahead of launch. It scored 41 on the Artificial Analysis Intelligence Index at high reasoning effort and 54.9% on HLE-Verified.
Why did GOOGL stock move after the report?
Alphabet shares rose about 0.7% in after-hours trading after a Wall Street Journal report said Google could release the model within days, reportedly backed by internal tests where Google engineers preferred it to Anthropic’s Opus in coding tasks.
Did Google confirm the Opus comparison?
No. That claim was attributed to people familiar with Google’s internal testing through its Jetski coding tool, not to an official Google statement or a published independent benchmark.
How does Gemini 3.8 Flash compare to GPT-5.6 Sol?
On the Artificial Analysis Intelligence Index, Gemini 3.8 Flash’s score of 59 at high reasoning effort matched GPT-5.6 Sol’s score at extra-high reasoning effort, according to a 192-run comparison test. Results varied across other individual benchmarks.
Is Gemini 3.8 Flash more expensive to run than Gemini 3.7 Flash?
The published per-token price stayed the same at $0.75 per million input tokens and $3.75 per million output tokens. However, because the model reasons more deeply by default, its effective cost per completed task was estimated at roughly 40% higher.
What is the codename “Skimaki”?
Skimaki was the internal project name reportedly used at Google before the model launched publicly as Gemini 3.8 Flash. Google has not publicly explained the name.
Where does Gemini 3.8 Flash rank on independent leaderboards?
It placed around 17th out of 196 models tracked on the HLE-Verified leaderboard as of October 2026, putting it in roughly the top 10% of evaluated systems, with output speed near 302 tokens per second.
Should developers switch from Gemini 3.7 Flash to 3.8 Flash?
It depends on workload. The model showed gains on structured, tool-using benchmarks but trailed rivals on long-horizon agentic tasks, so teams should test against their own use case rather than relying on a single aggregate score.