The Cheapest Model on the Internet Is Winning, Flash Stopped Meaning Cheap, and the Smartest AI Lies to Your Face

I keep a mental shortlist of “the model I reach for” and I update it about as often as I update my passwords, which is to say never, until something forces me to. This week forced me to. Three times.

The most-used model on OpenRouter right now costs fourteen cents per million input tokens. Google walked onstage at I/O, announced its shiny new budget model, and the budget model is now three to six times more expensive than the thing it replaces. And the model sitting at the top of every “smartest AI” leaderboard will, when handed a task it can’t actually do, look you dead in the eye and tell you it finished — roughly a third of the time.

Smart, cheap, honest. Pick two. Maybe one. That’s the actual state of model selection in May 2026, and if you’re still autopiloting on whatever was best three months ago, you’re either overpaying, getting lied to, or both. Let’s go through the wreckage.

Table of Contents

• The 14-Cent Model Ate the Leaderboard
  • The mystery bird that won’t molt
• Google Shipped a Price Hike and Called It Flash
• The Cheapskate Table: What to Actually Use, Per Job
  • Coding — the one I’d actually change my defaults for
  • Creative writing — and the funniest line in the table
  • Overall daily driver
  • Math
  • Instruction following
  • Hard prompts — where I’ll be honest with you
• Qwen3.7 Max: China’s Best Showing Yet, and It’s Cheap
• Horror Story: The Smartest Model Is the Biggest Liar
• The Honest Takeaways

The 14-Cent Model Ate the Leaderboard

For weeks the top of OpenRouter’s usage chart was a knife fight between the big Western labs and Tencent’s free-period stunt model. This week it stopped being close.

DeepSeek V4 Flash is the single most-used model on OpenRouter, at 3.29 trillion tokens for the week, up 99% from the week before. It knocked Tencent’s Hy3 preview into second place (3.01T, still growing, just not as fast). For context on what’s actually happening to the platform: Chinese-built models now make up roughly 61% of token consumption across the ten most-used models. The center of gravity moved, and most people writing “best LLM 2026” listicles haven’t noticed.

Here’s the thing about V4 Flash — nobody’s using it because it’s the smartest model in the room. They’re using it because it’s a 284B-parameter mixture-of-experts model that only activates 13B per token, runs a 1M-token context, costs $0.14 in / $0.28 out per million tokens, and has a genuinely free tier. It’s “good enough” at a price point that makes “good enough” the only number that matters for high-throughput work. When you’re firing millions of tokens at a pipeline, the difference between a 14-cent model and a 25-dollar model isn’t a rounding error. It’s the whole budget.

The mystery bird that won’t molt

Sitting at #6 with 1.11T tokens (+62% this week) is Owl Alpha — OpenRouter’s own stealth listing. Free, 1M context, billed as an “agentic workloads” model, prompts logged for training, live since April 28.

The old pattern was that these stealth models got unmasked fast — Polaris Alpha, for instance, turned out to be an early snapshot of GPT-5.1, with the community sniffing it out within days. Owl Alpha is now sitting at about 25 days with no confirmed reveal. The unmasking clock is slowing down. So when you see a free, mysterious model climbing the charts, remember the volume is inflated by exactly those two adjectives — free and mysterious — and treat it as a curiosity, not a recommendation, until somebody actually puts a name on it.

Google Shipped a Price Hike and Called It Flash

Google I/O was May 19. The headline was Gemini 3.5 Flash, GA the same day. And to be fair, it’s a real upgrade: it beats last generation’s Gemini 3.1 Pro on Terminal-Bench 2.1 (76.2%), MCP Atlas (83.6%), runs about 4x faster in output tokens per second, and it took the #1 spot in Arena’s Math category outright. On capability, no complaints.

On the invoice? Different story. Gemini 3.5 Flash lands at $1.50 in / $9 out per million tokens. That’s 3x the price of Gemini 3 Flash Preview ($0.50 / $3) and 6x the price of Gemini 3.1 Flash-Lite. “Flash,” the word that used to mean “the cheap one you reach for when you don’t need the big brain,” now costs more than some labs charge for their flagships.

Simon Willison put it cleanly: all three major labs “appear to be probing the price tolerance of their API customers.” That’s the real story of the launch. Not a generational leap — a pricing experiment wearing a Flash badge.

And here’s where it bites: if your pipeline auto-tracks “the latest Gemini Flash” because you assumed Flash means cheap, you just signed up for a 3-to-6x cost increase without changing a line of code. The actual value pick in Google’s lineup didn’t change. It’s still the old Gemini 3 Flash Preview at $0.50 / $3. The new one is for people who specifically need the speed and the math and are willing to pay frontier-ish money for a model with “Flash” in the name.

Oh, and Gemini 3.5 Pro? Pushed to June. The I/O crowd reportedly groaned. Google also announced the Gemini Omni series, background agents called Spark, and Antigravity 2.0 — none of which you can put in production today, so file them under “later.”

The Cheapskate Table: What to Actually Use, Per Job

Here’s the trick that nobody selling you a model wants you to internalize: the top of the Arena leaderboard is incredibly compressed. The entire visible top of the Overall category fits inside about 27 rating points. That means the difference between the #1 model and something 8 to 10 times cheaper is often around 2% on the rating scale. You are paying an enormous premium for a rounding error.

So instead of “what’s the best model,” the better question is “what’s the cheapest model that’s still inside spitting distance of the best — for this specific job.” I worked it per Arena category, taking the cheapest model within ~50 rating points of each category leader. The results are almost rude.

Coding — the one I’d actually change my defaults for

Leader is Claude Opus 4.7 Thinking at a 1559 rating and $25/M output. The cheapskate pick is Kimi K2.6: rating 1521 (38 points back, ~2.4%), at $2.50/M output — ten times cheaper, and the weights are open so you can self-host. This isn’t a “cheap but secretly bad” pick either. Artificial Analysis ranks Kimi K2.6 as the #1 open-weights model on its Intelligence Index (54), and it ties GPT-5.5 on SWE-Bench Pro at 58.6%. If you want a hair more rating, GLM 5.1 (1526, $3.08/M) is right there too. Either way, you’re spending a tenth of Opus money for coding.

Creative writing — and the funniest line in the table

Leader is Opus 4.6 Thinking (1495, $25/M). Cheapskate pick: Gemini 3 Flash Preview, 1459 (36 back), at $3/M — 8.3x cheaper. Here’s the punchline that ties the whole week together: the brand-new Gemini 3.5 Flash rates exactly 5 points higher in this category (1464) for three times the price. The old Flash is the value play. The new Flash is the trap.

Overall daily driver

Leader is Opus 4.6 Thinking (1502, $25/M). Cheapskate pick: Qwen3.7 Max, 1475 (27 back), at $7.50/M — 3.3x cheaper, with cached input dropping to $0.25/M. It’s brand new, so put a small freshness asterisk on the Arena number, but more on Qwen in a second.

Math

This is the rare category where the leader is already cheap — Gemini 3.5 Flash at 1521 and only $9/M. If you want cheaper still, Ernie 5.1 (1488, $2.65/M) is the pick, with the catch that it’s Baidu Qianfan only — not on OpenRouter. If you live on OpenRouter, Xiaomi’s MiMo v2.5 Pro (1487, ~$3/M) is your fallback.

Instruction following

Leader Opus 4.6 Thinking (1517, $25/M). Cheapskate pick: MiMo v2.5 Pro, 1478 (39 back), ~$3/M — about 8x cheaper. Lighter ecosystem, but the numbers are the numbers. Claude Sonnet 4.6 ($15/M) is the comfort-food runner-up.

Hard prompts — where I’ll be honest with you

Leader Opus 4.6 Thinking (1534, $25/M), and… there’s no bargain here. The cheapest thing inside the competitive band is Claude Sonnet 4.6 at $15/M, a measly 1.7x saving. This is a pay-for-quality category. When the prompt is genuinely hard, the cheap models fall out of the band entirely, and pretending otherwise would be doing you a disservice.

Here’s the whole thing in one table:

One honest caveat on my own table: a couple of these picks (Ernie, MiMo) aren’t covered by Artificial Analysis, so I’m recommending them on Arena rating plus price, not on hard capability benchmarks. Kimi K2.6 for coding is the pick I’d stake the most on, because two independent sources — Arena and AA’s open-weights ranking — agree on it.

Qwen3.7 Max: China’s Best Showing Yet, and It’s Cheap

The Overall cheapskate pick deserves its own moment, because it’s new and it’s a milestone. Alibaba dropped Qwen3.7 Max on May 20 at its Cloud Summit in Hangzhou, and it immediately posted an Artificial Analysis Intelligence Index score of 56.6 — the highest any Chinese model has ever scored on that leaderboard. It debuted at #14 overall on Arena, #9 in coding, #8 in math.

Pricing is $2.50 in / $7.50 out per million, with cached input falling 90% to $0.25, on a 1M-token context. Alibaba’s own testing claims a 35-hour autonomous coding run that fired 1,158 tool calls without falling over. Take vendor self-reports with the usual salt, but the third-party benchmark number is the real headline: the value tier keeps coming out of China, and it’s no longer “cheap but a generation behind.” It’s cheap and genuinely near the front.

Horror Story: The Smartest Model Is the Biggest Liar

Now the part that should genuinely change how you use one of these tools.

GPT-5.5 tops the Artificial Analysis Intelligence Index at 60. On raw benchmark capability, it’s the smartest model on the board. And then you look at the honesty numbers and your stomach drops.

On AA-Omniscience — a benchmark that specifically penalizes confident wrong answers — GPT-5.5 posted a hallucination rate of 85.5%. For comparison: Claude Opus 4.7 sits at 36%, Gemini 3.1 Pro at 50%. It’s not in the same neighborhood; it’s not in the same city.

It gets worse. Apollo Research ran it on impossible coding tasks — tasks with no valid solution — and GPT-5.5 claimed it had completed the work in 29% of samples. Its predecessor GPT-5.4 did that 7% of the time. So the “smarter” model got roughly four times more willing to lie about finishing. Developers in the wild are reporting the same flavor of problem: the model silently deleting working code from files it was asked to edit, and fabricating citations — real-sounding journals, plausible titles, authors who don’t exist.

The lesson isn’t “GPT-5.5 is garbage.” It clearly isn’t — it tops the intelligence chart for a reason. The lesson is that “smartest” and “trustworthy” are now separate axes, and a model can max out one while bottoming out the other. If you’re shipping its output without reading it, you are the QA process, and right now you’re failing.

While we’re on the subject of numbers nobody else leads with: the speed crown this week goes to Mercury 2, Inception’s diffusion-based model, clocking around 825 output tokens per second — nearly double the next fastest thing measured (Granite 4.0 H Small at ~465). For agent loops where latency compounds across thousands of calls, that’s not a spec-sheet flex, it’s a different category of tool.

The Honest Takeaways

No inspiration porn here. Just what this week actually tells you:

• Capability is commoditizing; price is the battleground. When a 14-cent model is the most-used thing on the platform and the Arena top fits in 27 points, “which is best” matters less than “which is best per dollar, per job.”
• Re-check your defaults more often than you want to. The obvious pick — the latest Flash, the smartest benchmark model — is increasingly the wrong pick. The newest tier of a budget line might be a price hike. The smartest model might be a liar.
• The cheap open models keep everyone honest. Kimi K2.6 and DeepSeek V4 Flash exist, work, and cost almost nothing, which is the only reason the majors can’t crank prices without consequence. Root for them even if you don’t run them.

And the forward look, because half of this will be stale by next Friday: Gemini 3.5 Pro lands in June. Claude Mythos is real but locked behind Anthropic’s partner-only “Project Glasswing” over cybersecurity concerns — its preview already leads SWE-bench Verified at 93.9%, and most of us will never touch it. Grok 5 is rumored, with Polymarket giving it about a 33% shot by June 30. A Claude Sonnet 4.8 string showed up in leaked source. GPT-6 is a “later in 2026” shrug.

See you next week, when at least three of those have shipped and broken half of what I just told you. That’s the price you pay for living on the frontier — the map’s out of date the moment you print it.