Same model, five effort settings. Artificial Analysis ran Claude Fable 5.2 through all of them:
→ low: 54, 5.4 min per task
→ medium: 56, 6.8 min
→ high: 62, 14.6 min
→ xhigh: 62, 25.7 min
→ max: 65, 40.8 min
xhigh buys zero points over high for 76% more time. Max buys 3 points for 2.8x the time of high.
Time to first answer: 24s at high, 212s at max.
Caveat: every Fable row on AA is labeled "Default Fallback", so requests its safety classifiers flag go to an older model and the score measures that whole setup.
Default worth trying: high. Save max for jobs you can leave running.
3
2
79
Stop judging a coding agent harness by its benchmark score. Judge it by its preamble.
A Shandong University study (arXiv 2610.04433) ran the same models through Claude Code, mini-SWE-agent and OpenCode on SWE-bench Verified:
→ pass rate: Claude Code vs mini-SWE-agent within 5 points on 447 tasks
→ swapping the harness flipped 13% of hard tasks, the same as just rerunning
→ cost per task: up to 3x apart
The bill is set before any work starts. System prompt plus tool schemas, resent on every step:
→ Claude Code: 16,581 tokens
→ OpenCode: 7,025
→ mini-SWE-agent: 829
What to do:
1. Log one raw request from your agent and count the tokens before your task text
2. Multiply by your average steps per task
3. Drop tools you never call and make sure that prefix hits the prompt cache
Caveat from the paper: cached pricing shrinks the bill but doesn't change the ranking.
6
7
213
People on X keep saying “nobody really uses Perplexity” while ChatGPT owns the timeline.
Meanwhile:
ARR just crossed $1B (CEO confirmed yesterday) up from ~$250M at the start of 2026 and $750M in August
100M+ monthly users across products
1B+ queries/month
Valuation: $20B last confirmed round (Sept 2025) → $30B+ talks with Nvidia still circling
IPO possibly as early as Q1 2027
The most efficient AI answer engine is scaling while everyone argues about the loudest chatbot.
4
2
51
Market share of AI referral traffic is smaller than ChatGPT/Gemini (often 5-15% depending on the study), and web traffic has plateaued relative to the hype cycle. That’s the point.
Perplexity doesn’t need to win the consumer attention war. It wins the answer + citation + agent workflow war, which monetizes better.
1
12
Users tell a different story than social media sentiment:
Core answer engine: ~45M MAU
All products (search + agents + Comet browser + API): 100M+
High retention (85-90% return within 30 days in multiple reports)
India alone spiked hard via the Airtel free Pro deal
It’s not “everyone is talking about it on X.” It’s “power users quietly switched their research workflow.”
1
16
Revenue growth is absurd:
June 2025: ~$150M ARR
March 2026: ~$450M
Aug 2026: $750M
Nov 2026: $1B+
That’s roughly 6–7× in ~18 months, with lower burn than the frontier labs. Computer (agents) + usage-based pricing did a lot of the heavy lifting.
1
16
Stop paying a chat model to answer yes or no.
Routing, policy checks, "is this ticket urgent", "which tool fits": most agent pipelines send these to a generative model and then parse the text it writes.
Google's new Decisions API (Nov 11) is built for exactly that job:
→ $0.10 per 1M input tokens, $0 for output
→ 150-250 ms end to end
→ up to 1,000 questions per call over one shared input, billed once
→ every answer comes back with a 0-1 confidence score
The pattern works with any provider: ask narrow yes/no or pick-one questions, branch on the probability in code, and only send low-confidence cases to the big model.
One action: list every place your agent asks an LLM a question with a fixed set of answers. Those are your cheapest wins.
8
5
239
GR00T 2's headline is 95% with five demos. The number I'd watch is 40%.
NVIDIA's new robot model (CoRL, Nov 9) is a world action model: it predicts future video and the robot's actions together, so it learns what the scene should look like after it moves, not just which move to make.
Pretraining mix:
→ 402,730 hours of first-person video of people at work
→ 10,184 hours of handheld gripper demos
→ 23,500 hours of robot teleoperation
On 50 tasks the test robot never practiced:
→ 40% success zero-shot, with 40 of 50 tasks succeeding at least once
→ 71% with one demo per task, 95% with five
→ five demos is under 3 min of robot data and a 30 min LoRA fine-tune on one RTX 5090
The catch: the whitepaper says the 50 tasks were picked from ones the model learns with five episodes, so the 95% is high partly by construction. Code is listed as coming in December.
The mechanism: predicting the scene, not just the action, is what lets human video teach a robot.
1
1
82
The "NVIDIA and OpenAI made agents 14x faster" story is really a KV cache story.
The paper is Evolving Infra (arXiv, Nov 2026), with authors from Georgia Tech, UIUC, Berkeley, NVIDIA, OpenAI and MIT.
The problem: serving stacks split GPUs into prefill nodes and decode nodes, and each node's KV cache is tied to its role. Agents keep re-sending long contexts, so when the workload mix shifts you either relaunch nodes or throw away cache the agent was about to reuse.
The fix: decouple the cache from the role.
→ A node switches between prefill and decode in 6.2s on average, vs 754s for a Dynamo relaunch
→ Lost KV reuse drops 81-95%
→ Up to 14x lower P95 time to first token, up to 4.1x lower mean latency, up to 2.7x throughput vs Dynamo
The 14x is a tail-latency number at the best setting, not 14x faster agents across the board.
The mechanism: for agents, the cache you keep matters as much as the GPUs you add.
Same prompt, same Blender 5.3: a mechanical watch that assembles itself, built in code by AI agents. What each run cost:
→ DeepSeek V4.1 Flash: $0.80 ($0 locally)
→ Kimi K3.1 Pro: $21.69
→ Claude Sonnet 5.5: $44.24
→ Claude Opus 5.5: $73.48
→ Claude Fable 5.2: $119.02
That's ~149x between the cheapest and the most expensive, and Fable at ~1.6x Opus.
Not my run, it's @ivanfioravanti's public test. His note: the models shared machines, so the times (3h7m to 8h30m) aren't like-for-like. Cost is tokens at API prices.
Look at the renders before the price tags. If the $0.80 watch is good enough for your use, that's the number that matters.
6
5
236
Reference
Quick update!
I've finally found a very tough one for Open Models here.
- DeepSeek V4.1 Flash ⏱ 3 hr 16 min 💰 $0.80 ($0 local)
- Kimi K3.1 Pro ⏱ 3 hr 7 min 💰 $21.69
- Claude Sonnet 5.5 ⏱ 5 hr 46 min 💰 $44.24
- Claude Opus 5.5 ⏱ 8 hr 30 min 💰 $73.48
- Claude Fable 5.2 ⏱ 5 hr 29 min 💰 $119.02 👀👀👀
- Claude Haiku 5.5 ⏳ final render
- Qwen4 Max ⏳ just started
- Hy4 planned
I will share videos and Blender models as soon as they are ready.
1
31
The chart behind Codex's "model at capacity" fix doesn't show more GPUs. It shows six retry releases.
Tibo's chart tracks the share of Codex daily users with at least one turn ending in a capacity error:
→ 6.82% at the Oct 19 peak
→ 0.085% by Nov 12
The six markers, Oct 15 to Nov 6, are all about when to retry:
→ WebSocket retry fix in the desktop app
→ a delay on ordinary overloads
→ a retry time for service-tier overloads
→ Codex forwarding retry and stop instructions
→ a delay for NO_MORE_RETRY
→ the delay kept inside WebSocket errors
The mechanism: when every client retries at once, a short overload turns into a retry storm. Tell clients when to come back and the spike flattens.
The chart can't rule out lower demand too. But the line sits near zero since the Nov 6 release.
4
7
200
A $4,000 coding agent bill doesn't come from one bad prompt. It comes from multipliers stacking.
Marc Lou says MailCheap cost him ~$4,000 to build with Opus 5.5 on Cursor, his first unprofitable product. His breakdown in the replies:
→ Ultra plan already maxed out
→ Opus 5.5 Max at 2x speed, billed at API pricing
→ 10 agents running in parallel
Each one multiplies the others. (He later guessed closer to $3.5K was MailCheap itself.)
Before your next parallel run, set a monthly spend limit. Cursor's help page has the steps for each plan and warns that enforcement isn't instant, so usage can briefly go past the cap.
Set it below what you'd be fine losing, then pick which multiplier you actually need today.
7
7
226
his breakdown teapawt.org/marclou/status/2121241…
It cost me ~$4,000 to build MailCheap with Opus 5.5 on @cursor_ai so ironically, it's my first unprofitable product.
1
40














