I help companies run on autopilot in 7 days. co-founder @ code2b.co & yourtrained.ai
Athens
- Tweets 8,468
- Following 447
- Followers 562
- Likes 11,972
minimax m3.1 flash: $0.15 in / $0.60 out per million tokens. deepseek's off-peak rate.
artificial analysis: when it didn't answer right, 93% of the time it answered wrong instead of saying it didn't know.
cheap tokens. confident mistakes.
check it before it writes to a CRM.
1
1
36
a researcher found attackers' AI agent panels left open online. the agents tried 1,200+ company websites and got into sites at 100+ companies.
his estimate to reuters: ~$5,000 in tokens. roughly $50 per break-in.
breaking in got cheap. patch the old stuff first.
2
162
IBM survey, 3,221 execs at large firms in Europe and the Gulf: 80% say an AI incident reached their board or leadership last year.
Of those, 40% changed their AI governance policy.
An incident that changes nothing is a rehearsal for the next one.
1
3
135
cloudflare's new extraction model was asked for a tracking number that wasn't in the email.
it returned null. its best guess was "$52.10". the tax.
$0.10 per million input tokens. output is free.
for invoices and order emails, "not in here" is the most useful answer.
1
1
6
227
deepseek opened its own search API this week. $0.75 per 1,000 searches.
run deepseek inside claude code or codex? your agent's web searches already hit it. free for now.
billing starts nov 1.
at 10,000 searches a day that's ~$225 a month. count your agent's searches now.
2
2
6
220
DHS ordered US civilian agencies to remove DeepSeek and stop buying it.
The order also covers third-party products built using DeepSeek.
So the question isn't "do you use DeepSeek?"
It's "which models are inside the tools you already pay for?"
Could you answer that today?
1
3
121
François Chollet: agents write ~100% of the code, engineers have never worked harder, and the bottom line "is typically not budging."
Automate one step and the bottleneck moves.
Before you automate anything, name the business number it should move. Then check it moved.
2
103
google research open-sourced harness2. it tunes everything around the model: prompt, tools, memory, skills. weights untouched.
one step: up to +11.5 points on one benchmark, same model. their own scoring.
before paying for a bigger model, fix the setup around the one you have.
3
2
10
304
microsoft pulled out of a planned AI datacentre in essex. the grid can't power it until the early-to-mid 2030s.
the government called it the "largest UK sovereign AI datacentre".
GPUs you can buy. power you wait for.
planning on "local AI capacity"? ask for the grid date.
3
1
5
122
Survey of 850+ IT decision-makers: 82% say they can keep pace with governing AI agents.
26% feel well prepared to resolve conflicts between agents.
5% have automated controls, audit trails and human review points. Europe: 3%.
Governance is a PDF until it runs in the workflow.
2
1
3
142
deepseek's new flash model: $0.15 per million input tokens. that's the headline.
the footnote: that's off-peak. weekdays 06:00–10:00 UTC it's $0.30. output: $0.60 becomes $1.20.
that window is the european working morning.
run batch jobs after lunch. same model, half the bill.
1
1
1
110
Yelp now lets AI agents in only if they pay for its data licence. Delta is blocking agents from booking flights.
The web is splitting into "pay us" and "go away".
Founders: when a customer's AI agent tries to buy from you, what happens today? Do you even know?
7
1
11
307
Microsoft's Copilot Studio agents now get their own account: mailbox, Teams, Entra ID.
Read the memory section. One memory, shared with everyone it works with. Files you share can surface in answers to people without access. No opt-out in preview.
Scope what it sees first.
1
3
176
the test for a bulk upload isn't one good file.
it's 30 mixed ones. some too big, some the wrong type.
done means: the valid ones go through with visible progress, every rejected one says why, and failed ones can be retried.
demos upload one pdf. users drop a folder.
1
1
6
114
matt pocock: "I think I'm addicted to getting agents to build state machines."
not just for frontend.
a client's report this week said "1 overdue" up top. no overdue row below.
if a count and its list can drift apart, they will.
one state per item. the header counts the rows.
2
5
195
A healthcare billing firm's CPTO had tried AI. Verdict: it "did a bunch of silly crap."
Rebuild: seven agents, one workflow each. Plain code accepts or rejects every action. Off-allowlist output fails closed.
Main agent: 59 of 60 on evals.
The model suggests. Code decides.
2
1
2
71
elorian's new vision model lists tokens 94% cheaper than GPT-6 Astra.
in their own robot-planning demo, one episode cost $0.035 vs $0.249. that's 86% cheaper.
still great. but they say looking closely takes time and tokens.
compare cost per finished task, not per token.
1
1
3
119
mercury's MCP lets ChatGPT or Claude request a payment. a person approves in the app before money moves.
their docs: "A confirmation in chat is not approval" and "request created" does not mean "payment sent."
gate what you can't undo. never let "requested" read like "done".
1
38