In partnership with
FROM OUR PARTNER
100+ Claude Code hacks to ship code 10X faster
Top engineers at Anthropic and OpenAI say AI now writes 100% of their code.
If you're not using AI, you're spending 40 hours doing what they do in 4.
These 100+ Claude Code hacks fix that and help you ship 10x faster.
Sign up for The Code and get:
100+ Claude Code hacks used by top engineers — free
The Code newsletter — learn the latest AI tools, tips, and skills to code faster with AI in 5 minutes a day
TODAY AI BRIEF
3 numbers behind OpenAI's cheaper agent models
1 GPT-5.6 Luna nearly matches GPT-5.5 on BrowseComp at a fraction of the price.
The reported scores are 84.04% for Luna at $1.33 and 84.36% for GPT-5.5 at $33.27. BrowseComp is a search benchmark for hard-to-find facts, not a computer-use test, and these figures are company claims. Read StartupHub's breakdown.
2 Browser Use reports a near-equal result for a fraction of the cost.
Its test found Luna completed 78% of challenging browser tasks for about $14, against 80% for about $235 with the prior state-of-the-art model. That is a customer's own result, so the test conditions matter. See the customer figures.
3 The 'halves the cost' headline is not in the coverage we could open.
The Next Web reports that GPT-5.6 Sol uses 54% fewer output tokens, a figure OpenAI gives without a stated baseline, and that Luna costs 80% less than Sol without saying whether that is per token or per task. Read The Next Web's report.
AI FORWORD
Cheaper agents are real, but the savings depend on the test you run
The shift. Agents make many model calls for one job, so the price of a single answer matters less than the cost of a finished task. Vendors now compete on that cost. StartupHub's overview lists the per-task figures they report.
The evidence. The strongest numbers are the BrowseComp comparison and Browser Use's test, where a cheaper model came close to a pricier one. Both are self-reported, and the BrowseComp scores come from different reasoning settings, which makes them hard to compare directly. Check the caveats here.
The gap. Several baselines are missing. The Next Web gives a 54% token cut without its comparison point, and it says OpenAI has since cut prices again without publishing new figures. Read the coverage.
The reframe. Stop asking which model is cheapest per token. Ask what it costs to finish your task correctly, including the retries. That number decides whether an agent pays for itself, and vendor claims rarely give it to you.
PRESENTED BY :
Stop Making Shoppers Spell It Out
“Running shos” shouldn’t return zero results. Neither should “red waterprof trail runners under $150.”
Traditional search needs exact keywords. Gladly understands what shoppers mean, typos and all, and turns dead-end searches into products that fit what they asked for.
SOCIAL SNAPSHOT
What to check before you switch models
$1.33 against $33.27. That is the reported BrowseComp cost gap between Luna and GPT-5.5, with near-equal scores. Read the figures.
The reasoning settings differ. The comparison sets Luna at Extra High against GPT-5.5 at Extra High, but some other results use lower settings. Compare like with like. See the setup notes.
A reasoning jump. On ARC-AGI-3, GPT-5.6 Sol reportedly rose from 13.3% to 38.3% with retained reasoning and compaction, using about six times fewer output tokens. Company claim. Read the benchmark note.
Prices keep moving. Both articles mention price cuts after launch without new figures, so any cost table you build will go stale. Read the coverage.
PROMPT STATION
Price one agent task before you automate it
Most automation projects start with a model choice and skip the cost of failure. This prompt works in Claude, ChatGPT or Gemini and turns one task into a cost-per-completed-task estimate you can check before you build.
I want an AI agent to do this task: [DESCRIBE ONE TASK, STEP BY STEP, WITH THE TOOLS IT USES]. I will run it [NUMBER] times a month.
Act as a cost analyst. Estimate the calls and tokens per run, then ask me for the price of each model I plan to test. Show cost per run, then cost per completed task using a success rate of [SUCCESS RATE] and [RETRIES] retries per failure.
Finish with a 20-run test plan: which runs to include, how to score success, and the single number I should compare across models before I commit.Tip: Swap the brackets for a real task, such as sorting supplier invoices or checking product listings. Use your own success rate from a small trial, not a vendor's number, and rerun the estimate whenever prices change.