In partnership with
In Today’s Issue:
🧗 Opus 5.5 climbs to the top, GPT-6 gets much cheaper
📉 Epoch: the price of AI thinking falls 13x a year
🎓 a16z opens an academy with no degree
📱 Qwen puts AI agents on your phone
🧬 Claude makes biology's AI models 4x faster
✨ And more AI goodness…
⚡ The Signal
The new leader in AI rankings costs less to run than the model it replaces, and its biggest rival just halved the price of its everyday models.
Anthropic's Claude Opus 5.5 now tops Artificial Analysis's intelligence ranking at about 40% lower running costs than Opus 5, while OpenAI cut API prices for GPT-6 Sol and Luna in half. Epoch AI puts a number on the pattern behind both: the cost of reaching a given level of AI performance has fallen about 13-fold a year for three years, faster than electricity, batteries or DNA sequencing ever got cheaper. Cheaper thinking changes what is worth handing to a machine, from phone agents to the protein models Claude has made four times faster. And Washington has decided what to call all of this.
All the best,
Kim Isenberg
Relative cost of AI (2021-26) against DNA sequencing, lithium batteries, compute and electricity, each from the start of its price decline. Chart: Epoch AI.
📉 AI Thinking Is Getting Cheaper Faster Than Anything Before It
Epoch AI finds the cost of a given level of AI performance has fallen about 47% per quarter, or 13x a year, over the past three years, faster than any other transformative technology it compared. In January 2025, OpenAI's o3 reached a 75% score on the PhD-level science exam GPQA Diamond for about 30 cents per question; GPT-5.6 Luna matched it for $0.0004, a 725-fold drop in under 18 months. Prices fall fastest right after a capability first appears, about 75x a year, and slow to roughly 4.7x a year two years later.
👉 tl;dr: The same AI answer that cost 30 cents in early 2025 now costs a fraction of a cent, so more tasks keep becoming cheap enough to automate.
Announcement image for the Horowitz Andreessen Academy. Image: a16z.
🎓 a16z Opens an Academy With No Degree and No Homework
Andreessen Horowitz is launching the Horowitz Andreessen Academy, a free one-year program in San Francisco for 50 students, mostly recent high school graduates, starting in fall 2027. Backed by $42 million led by a16z, it partners with Anthropic, OpenAI, Google, Meta, Nvidia, Palantir and four others: students take short classes from figures like Sam Altman, work inside partner companies and get more than $50,000 in compute credits, but no degree or accreditation. It is the latest billionaire-built alternative to college, this time from founders who have openly criticized universities.
👉 tl;dr: A top venture firm is offering teenagers a year of hands-on work inside leading AI and tech companies as an alternative to college.
Qwen Intelligence launch graphic. Image: Alibaba Qwen.
📱 Qwen Wants an AI Agent Running Your Phone
Alibaba's Qwen team launched Qwen Intelligence, three AI agents built to get things done on a smartphone. A Planner breaks a request into steps, a Mobile-Use Agent carries them out across apps, using app interfaces where it can and tapping the screen where it must, with a reported 90% end-to-end success rate, and a Creative Agent turns one sentence into a finished image in about 3 seconds. Qwen is also opening its test suites, including a safety benchmark, and lists a standard subscription in China at 78 yuan a month.
👉 tl;dr: Qwen now sells a phone assistant that plans and completes tasks across your apps, and is publishing its tests so others can check the results.
Making AI video is usually a tool-juggling nightmare for me, jumping between single-purpose apps, only to end up uneditable renders. One typo shouldn't mean rebuilding an entire project from scratch.
Fotor Agent changes the equation by turning that entire workflow into a single. chat-driven process.
Instead of handing you a locked, black-box render, its long-horizon planner dynamically orchestrates video, audio, and AE-level 4K motion graphics directly onto a fully editable multi-track timeline.
You can adjust every element right on the canvas. from precise data charts and text overlays to semantic video editing that cuts filler and pairs B-roll automatically.
Fotor claims it delivers professional motion graphics at 50x the speed and 1/200th the cost of traditional After Effects workflows—marking the shift from a gimmicky AI demo to a production-ready tool you can actually rely on for real client work.
🎬 Watch This
In this 56-second showcase from Anthropic, Opus 5.5 traced 2,781 points along the Moon's horizon in Apollo 8's famous 1968 photo, matched them against lunar elevation data and worked out exactly where and when it was taken: 16:39:39 UTC on December 24, 1968, about 69 miles above the Moon. It then rebuilt the scene in 3D and compared it with the original, a compact look at the long, multi-step work Anthropic built Opus 5.5 for.
“GPT-6 Sol and Luna are big improvements on intelligence, alignment, work output, coding, computer use, and more over their 5.6-family predecessors. They are also half the price per token, and even less per task!”
– Sam Altman, CEO of OpenAI, September 22.
The closing phrase is the real claim: in OpenAI's own business-workflow tests, Luna scores higher than its predecessor at 58% lower cost per task.
Trump says the US will now call AI “super intelligence.”
At the UN General Assembly on Tuesday, he said the word “artificial” makes the technology “sound fake” and ordered the new term into all US documents, even though it was not an option in his own weekend poll, where Superior Intelligence won with 41.5% of about 234,000 votes.
Opus 5.5 Climbs to the Top, GPT-6 Gets Much Cheaper
The Takeaway
👉 New leader: Claude Opus 5.5 scores 58 on Artificial Analysis's Intelligence Index, ahead of Fable 5.1 and GPT-6 Astra at 53 each.
👉 Cheaper to run: Anthropic says Opus 5.5 costs about 40% less than Opus 5 on typical work, at $4/$20 per million tokens.
👉 Half price at OpenAI: GPT-6 Sol drops to $2/$10 and Luna to $0.10/$0.50, both trained with GPT-6 Astra's methods.
👉 Tighter guardrails: Opus 5.5 hands most cybersecurity tasks to an older model; vetted labs can apply for fuller biology access.
The AI frontier moved in two directions on Tuesday: Anthropic raised the ceiling, and OpenAI cut the price of getting close to it. Claude Opus 5.5, the first model in Anthropic's new Claude 5.5 family, performs at the level of the company's flagship Fable 5.1 on most work, according to Anthropic, and now leads Artificial Analysis's independent intelligence ranking. Later the same day, OpenAI brought the training methods behind its top model, GPT-6 Astra, to the cheaper GPT-6 Sol and Luna, and cut their API prices in half.
Opus 5.5 is built for long, sprawling jobs. One early tester completed a 680,000-line code migration in less than a day. In an internal test, it rewrote HAProxy, widely used software that spreads web traffic across servers, from C into Rust in 9.5 hours, against 12 for Fable 5.1, at 51% lower cost. Anthropic says the savings come from a lower price per token and fewer tokens per task. It also addressed one of the most common complaints about Opus 5: the new model writes more plainly and puts the most important information first. On Terminal-Bench 4.0, a test of multi-step work in a command line, Anthropic reports 66.4%, ahead of the 57.9% OpenAI reported for GPT-6 Astra.
Terminal-Bench 4.0 scores as reported by Anthropic; Opus 5.5 at xhigh effort, GPT-6 Astra and GPT-5.6 Sol as reported by OpenAI. Graphic: Superintelligence; data: Anthropic.
OpenAI's pitch is near-top intelligence for much less money. GPT-6 Sol now lists at $2 per million input tokens and $10 per million output tokens, half of GPT-5.6 Sol's promotional pricing, and Luna falls to $0.10 and $0.50. On OpenAI's internal factuality test, Sol makes about half as many mistakes as its predecessor, and on the business-workflow test AutomationBench Sol at its xhigh setting beats Claude Opus 5 at max effort for 9% of Opus 5's cost per task. Both models are rolling out in ChatGPT Work and Codex for paid plans, with Luna also available to free users in the desktop app.
API list prices per million tokens before and after the September 22 launches; OpenAI compares with GPT-5.6 promotional pricing. Graphic: Superintelligence; data: Anthropic, OpenAI.
Most of these numbers come from the companies themselves, and Anthropic concedes that at this level “benchmark margins have become a less reliable guide to real-world differences.” Its safeguards also shape the scores: when they block a cybersecurity or biology task, an older Claude model finishes it. Still, the direction is unmistakable, and it is the same one Epoch AI measures in today's news: the price of a given level of AI performance falls fastest right after that level first appears. This week, the new leader got cheaper to run, and OpenAI's everyday models got half as expensive.
Why it matters: Both leading labs made top-tier AI cheaper on the same day. For anyone running agents on long coding or office work, handing over bigger jobs just became much more affordable, whether they choose the new leader or OpenAI's discounted tier.
The chart: Artificial Analysis's Intelligence Index v4.3, a combined score across 10 hard evaluations from coding to Humanity's Last Exam, puts Claude Opus 5.5 at 58, five points clear of Claude Fable 5.1 and GPT-6 Astra (53 each) and seven above Opus 5 (51). The lower panel plots the same index against the average cost per task, on a log scale.
The lesson: Anthropic did not buy the jump with a bigger bill. From about $1 per task upward, Opus 5.5's settings form the chart's best-value line, and a cheaper setting already reaches Opus 5's max-effort score for roughly a fifth of the cost. Its top score costs about $6 a task.
The caveat: The Opus 5.5 and Fable 5.1 bars are labeled “with fallback”: when safeguards block a cybersecurity or biology task, another Claude model finishes it. And GPT-6 Sol and Luna, launched the same day, are not on this chart yet; the Sol and Luna shown are the older GPT-5.6 versions.
🧬 Claude Made Biology's AI Models Four Times Faster
⚡ Bottom line
Claude sped up more than 30 open-source biology models about fourfold in under four weeks, and Anthropic is releasing the code.
💡 Why it matters
Protein design that used up to $10,000 of computing per target now reaches similar simulated scores for about $150.
🔎 What it means
Cheaper, faster tools let smaller labs run work that used to need a well-funded team.
The software scientists use to predict and design proteins, a key step in drug discovery, now runs about four times faster, and Claude did the engineering. In a research post published on September 17, Anthropic describes how. Supervised by two staff members who knew biology but had never worked on speeding up code, Claude optimized more than 30 open-source models in just under four weeks, work that usually takes experienced engineers weeks for each model.
The first payoff is speed. Across more than a dozen structure-prediction models, the programs that work out a protein's 3D shape from its sequence, Claude's changes made them run about 4x faster on average with minimal loss of precision, and about 1.6x faster with identical outputs.
Speed-ups on protein structure models; the right panel shows the averages: 1.6x with identical outputs, 4.2x in fast mode, 3.5x in Big mode. Chart: Anthropic (cropped).
The second is size. Claude built a low-memory “Big” mode that predicts molecular machines made of more than 10,000 building blocks, such as the bacterial ribosome, the cell's protein factory, on a single GPU server. Those predictions closely matched structures measured in the lab. At more than 31,000 building blocks the predictions fell apart, far beyond anything the models were trained on.
Large molecular machines predicted with Big mode (colored) over their experimental structures (gray); the bottom row shows less accurate predictions. Image: Anthropic.
Then the cost. In Anthropic's August protein-design work, Claude could spend up to $10,000 of computing per target to design proteins that latch onto a chosen molecule. With the faster tools, one Claude model on one GPU for 24 hours reached comparable simulated binding scores for about $150 in computing and tokens. Simulations are not lab results, so Anthropic and Adaptyv Bio are backing a competition that will test more than 5,000 designs in a real lab, with up to $1 million in Claude credits. All the optimized code is open source, so any lab can run the faster versions today.
The best in influencer marketing. And you’re invited.
Get ready for Return on Influence Festival '26. An entire day dedicated to influencer marketing with speakers running some of the best programs in the world. October 21, 2026, 10 am - 5 pm ET.
