Gemini 3.8 Flash shipped on September 2, 2026, Google's third Flash model in six weeks, together with Gemini 3.8 Flash Cyber, a version tuned to find and patch vulnerabilities that you cannot buy. Four hours later Meta shipped Muse Spark 1.3 with a price tier that is twenty times cheaper if Meta may train on your sessions. If you pay for tokens, all three change the maths you do before picking a model, and the per-token price is the least useful number in either announcement.
contributor endpoint costs $0.10 in and $0.20 out, and the only difference is that Meta uses your data to improve its products.
Google's announcement calls it "our best reasoning and coding model yet, at the same speed and low cost of 3.7". Gemini 3.7 Flash came out three weeks earlier. Logan Kilpatrick counted "our 3rd updated Flash model in only 6 weeks".
The headline numbers are Google's own: 54.9 % on HLE-Verified, wins claimed on Vals Finance Agent V2 and Harvey's Legal Agent Benchmark, and on DeepSWE v1.1 "3.8 Flash outperforms most larger frontier models". Kilpatrick posted 73.7 % on DeepSWE. The account Chubby wrote that "Flash outperforms 5.6 sol and opus 5 on terminal bench 2.1, HLE and much more."
The Hacker News thread reached 1,157 points. The top comment was Simon Willison's quick test: "make me a cool thing in html" returned a particle simulation in 13 seconds for 1.8 cents. Another commenter noticed that its "60 FPS" counter was hard-coded into the page.
The pricing sits in a footnote of the launch post: "Introductory price expires on December 31, 2026. Starting January 1, 2027, $1.50/1M input tokens and $7.50/1M output tokens will apply."
| Gemini 3.8 Flash | Input / 1M | Output / 1M |
|---|---|---|
| Until Dec 31, 2026 | $0.75 | $3.75 |
| From Jan 1, 2027 | $1.50 | $7.50 |
So "the same low cost of 3.7" holds for four months. HN user hiddencost put the problem plainly: "you're effectively planning to charge users twice as much for a model that is no longer frontier." Google keeps 3.7 Flash "fully supported for efficiency-first workloads", which is your fallback if the January bill matters.
Google's post explains the gains in one sentence worth reading twice: "3.8 Flash works harder." It works harder by "executing extra reasoning steps, and calling tools iteratively. At times, the model might use more tokens to maximize performance."
That is a design choice, and the independent numbers confirm it. The DeepSWE v1.1 leaderboard (Datacurve, 113 tasks, the same mini-swe-agent harness for every model) shows:
| Model | Pass rate | Avg cost / task | Output tokens / task | Steps |
|---|---|---|---|---|
| gemini-3.8-flash [high] | 74 % | $2.36 | 143k | 166 |
| claude-opus-5 [max] | 74 % | $11.84 | 118k | 99 |
| gpt-5.6-sol [max] | 73 % | $6.46 | 60k | 61 |
| gemini-3.7-flash [medium] | 65 % | $2.03 | n/a | n/a |
Read it both ways. Flash ties Opus 5 for about a fifth of the cost per task, which is the real result of the day. It also takes 2.7 times the steps and 2.4 times the output tokens of Sol. Cheap per token, chatty per task.
Artificial Analysis saw the same thing from another angle. Its Intelligence Index puts Flash at 47, 28th of 202 models, and running that index took 140 million output tokens against a median of 79 million. Grading it cost $1,077.95. Output speed is fast, 280.8 tokens per second, but time to first token is 12.74 seconds against a median of 3.33, because the model thinks before it speaks.
What this means if you run agents: the number to track is cost per finished task, not price per token. A simplified sketch, using only the DeepSWE output-token average and list prices (input tokens excluded, so the real figure is higher):
out_tokens = 143_000
print(out_tokens / 1e6 * 3.75) # 0.536 USD per task today
print(out_tokens / 1e6 * 7.50) # 1.0725 USD per task from Jan 1, 2027
The twelve-second first token hurts chat UIs more than background agents.
The more interesting half of the launch is Gemini 3.8 Flash Cyber: the same model with its safety limits loosened for what Google calls trusted defenders. Sundar Pichai posted that it achieves "86.2% on the important CyberGym industry" benchmark, which measures autonomous vulnerability discovery.
Google's post adds the rest:
Google says it "prioritized [fixing] over offensive capabilities like exploitation". Then it gates the model anyway. Access goes through the new Fairwind Program, which has "more than 650 participating partners globally", is limited to internal security, incident-response and pentest teams with multi-factor authentication, and pairs the model with Google's CodeMender harness.
The contradiction is the point. If a model only fixes, you can sell it to anyone. Gating it tells you Google thinks it can also walk through the holes it finds. Every frontier lab now has a security model it will not sell to the public; this one is cheap to run.
That evening Meta published Muse Spark 1.3 (HN, 690 points). Mark Zuckerberg called it "frontier performance almost too cheap to meter". That is true for one of the two endpoints on Meta's pricing table:
| Endpoint | Data use | Input / 1M | Cached / 1M | Output / 1M |
|---|---|---|---|---|
muse-spark-1.3 |
"Not used to improve our products" | $1.25 | $0.15 | $4.25 |
muse-spark-1.3-contributor |
"Used to improve our products" | $0.10 | $0.002 | $0.20 |
The private endpoint is more expensive than Gemini 3.8 Flash. The contributor endpoint is about 12.5 times cheaper on input and about 21 times cheaper on output, and you pay the difference with your data. Meta did not hide it in a license; it wrote the license as a price list. On HN, jmward01 said so: "it is now completely obvious how much stealing my tokens for training is worth to model providers … This is the first quantifiable number I have seen." HDBaseT asked the follow-up every security team will ask: "whether anyone has yet extracted AWS keys from a model trained on user input."
On performance, Meta's own table gives Muse Spark 1.3 (max) 75.4 on DeepSWE v1.1, above Opus 5 at 74.0 and Sol at 73.0, up from 55.0 for Spark 1.2. That is Meta's table, not Datacurve's board. Meta's engineers say 1.3 uses about 20 % fewer tool calls and 25 % fewer tokens than 1.2, the opposite bet from Google's "works harder".
215,128 pages written for Perplexity. Trellner Research asked Perplexity's Sonar models for the best software in 380 categories and read all 7,534 citations. 59.8 % pointed at domains ranked below 100,000 on Tranco, 23.4 % at domains outside the top million, and Wikipedia was cited three times. Three sister sites on the same Cloudflare nameservers had published 215,128 generated /best/<x>-software/ pages; two title their homepage "Facts & Grounding Page", which is addressed to a crawler, not to a person. (HN)
Hang on to your Firefox. Mark Rogers argued that "Firefox is our last best hope for browser engine diversity" and got 988 points on HN.
Mistral trains on Vibe by default. Mistral's help page says Vibe users "are not opted out by default"; Enterprise is, and the Vibe and API toggles are separate. On HN maxdo summed it up: "It's a spyware, but a sovereign one".
I stamped Gemini 3.8 Flash SHIP IT. It ties Opus 5 on DeepSWE, a board neither Google nor Meta runs, for about a fifth of the cost per task. The small print: it spends tokens freely, thinks for twelve seconds before answering, and doubles in price in January. Read the token bill, and read Meta's column headers.
How much does Gemini 3.8 Flash cost?
$0.75 per million input tokens and $3.75 per million output tokens until December 31, 2026, then $1.50 and $7.50.
Can I use Gemini 3.8 Flash Cyber?
Only through Google's Fairwind Program, which is limited to governments, critical-infrastructure operators and vetted security teams using multi-factor authentication.
Is Gemini 3.8 Flash better than Claude Opus 5?
On DeepSWE v1.1 they tie at 74 %, with Flash at $2.36 per task and Opus 5 at $11.84. Flash uses more steps and output tokens to get there.
This article expands on an episode of The Daily Diff, a five-minute daily video on what shipped and what broke in tech. Watch the episode · Subscribe on YouTube · the written diff lands in your inbox every morning at thedailydiff.dev.