gpt-5 on S Anand

LLM Deprecations and Price Changes

Thu, 21 May 2026 14:02:10 +0800

A colleague told me a near-miss horror story.

As Google began deprecating Gemini 2.0, we moved to Gemini 2.5 Pro. But reasoning is enabled by default and cannot be turned off.

For our specific problem statement, reasoning was not required. Token costs increased 10x and speeds were 3-4x slower. We moved the client to Gemini 2.5 Flash Lite, which has reasoning turned off by default and offers much lower latency.

Because we track compute costs closely, we managed this without a major financial impact.

But model updates clearly require careful testing on the cost and latency front as well, not just output quality.

AI used to keep getting cheaper. But now it’s more a “convergence”.

Each line traces a model family. The X-axis is its intelligence (LMArena ELO score) and the Y-axis is the input cost ($ per million tokens, log scale). Time flows roughly left to right as models improve.

Three patterns emerge and each seems strategic.

Gemini Flash rockets upward (cheap -> expensive, moving right).
GPT-4 class collapses from $30 toward $1.25, then GPT-5.5 jumps back up to $5.
Claude Sonnet runs perfectly flat (cheap stays cheap as it gets smarter).

Google’s Loss Leaders

Gemini 1.5 Flash was $0.075/MTok in August 2024. I told everyone to use it - it’s a fantastic, affodable model.

Gemini 3.5 Flash, released this week at Google I/O, is $1.50. That’s a 20× in 21 months. Looks like “Flash” is more a brand position than a price position. It migrates up-market.

Same for Gemini 1.5 Flash 8b (3.75 cents) which migrated into Gemini 3.1 Flash Lite (25 cents - a 6.7× increase).

Gemini Pro went the opposite direction: down from $5.00 (1.5-pro in Oct 2024) to $1.25-$2.00 today. Pro seems to be a competitive weapon against Anthropic and OpenAI at the top, while monetizing the middle.

Of course, Gemini’s real lock-in is the Google Workspace bundling and Search AI Mode. Personally, I subscribed to Google Pro for the first time in 20 years just for these bundled capabilities.

By the time people notice the Flash price, it’s hard to leave the ecosystem.

OpenAI’s Relaunches

The price chart speaks for itself:

Model	Price
GPT-4	$30
GPT-4 Turbo	$10
GPT-4o	$2.50
GPT-4.1	$2.00
GPT-5	$1.25
GPT-5.2	$1.75
GPT-5.4	$2.50
GPT-5.5	$5.00

Maybe their strategy seems is: scale current technology to commodity (the O-series showed the same pattern with $15 for O1 falling to $1.10 for O4-mini in 7 months), THEN launch a new frontier above it, and repeat?

Anthropic’s Revisions

Claude Sonnet has held at exactly $3.00/MTok input for over two years, across four model generations and an ELO gain of nearly 200 points. Pretty unusual.

Opus came down from $15 to $5 in November 2025 - likely a deliberate move to make it production-viable. Haiku crept from $0.25 to $0.80. The tiers are converging.

There’s a twist, though. Anthropic restructured enterprise contracts in late 2025. Seat prices dropped to $20/seat. Token discounts (previously 10-15% off API rates) were removed. For a 100-seat team, that adds ~$15K-$40K to annual TCO. So, prices went down, but the actual bill went up.

Also, Opus 4.7 uses a new tokenizer that may consume ~35% more tokens for the same text. It’s worth re-benchmarking prompts before assuming $5 is 1/3rd of $15.

What Do We Do?

Model family prices change rapidy. Old models get deprecated. Best to be prepared.

Add multi-tier routing to your architecture:
- Lite / Nano / Haiku for extraction, classification, … – tasks with clear right answers
- Sonnet / GPT-5.4 / Gemini Flash for most production reasoning
- Opus / GPT-5.5 for escalation or expert advice: complex planning, hard edge cases, high-value decisions
Compare completed-task costs, not token price. A 2* more expensive model can halve the retry rate, making it cheaper per successful output.
Migrate by model capability, not model family. Switch to models with similar latency, context window, output format compliance and reasoning depth.
Evaluate open-source models. DeepSeek models at self-hosted inference costs can be 90% cheaper for commodity (not frontier) tasks.

For enterprise procurement: keep a close eye on pricing changes, API token discounts, and what was quietly removed during renewal. (AI helps with this!)

The first 18 months of most AI model families are discounted customer acquisition. Then value extraction follows. Google started it with Flash. OpenAI is doing it with GPT-5.5. Anthropic is doing it with enterprise billing restructuring.

Fair enough. Providers need to recover infra investments.

And build good routing strategies helps enterprises get the most out of this.

Just keep asking yourself: “what’s our plan for when this model changes or deprecates?”

PS: AI-generated image - has a few errors.

Things I Learned - 30 Nov 2025

Sun, 30 Nov 2025 00:00:00 +0000

This week, I learned:

Warp has a terminal agent feature - allowing Warp to control a terminal via text. I find that regular coding agents like Codex can do that too with tmux. For example, I opened a session and had Codex run commands in it while I watched. Here’s the guidance it needed:

# Create a new session
tmux new-session -d -s $SESSION 'uv run --with pandas,httpx,lxml python -iqu'
# Capture output to a log file
tmux pipe-pane -t $SESSION -o "cat >> /tmp/$LOG"
# Run a command
tmux send-keys -t $SESSION 'print(1 + 2)' C-m
# See output
cat /tmp/$LOG
# Capture the last 5 lines of the pane
tmux capture-pane -p -t $SESSION -S -5

Notes from Early science acceleration experiments with GPT-5 - via Claude
- LLMs are accelrating research because they are good at:
  - Literature search, especially across disciplinary boundaries
  - Generating and checking routine calculations
  - Proposing variations on known techniques
  - Identifying connections between disparate results
  - Producing first-draft code for well-specified problems
  - Explaining why certain approaches won’t work
- But they’re curently struggling with the following - though it’s a shrinking space
  - Genuinely novel conceptual leaps (but this is increasingly happening, e.g. Sawhney and Sellke’s problem #848)
  - Recognizing when it’s plagiarizing, e.g. when it “discovered” a proof for the Chevalley-Warning theorem which was copied from a Noga Alon paper - it wasn’t conscious of this
  - Knowing what it doesn’t know
  - Distinguishing important problems from unimportant ones
  - Understanding the “negative space” of mathematics (why certain problems are hard, why obvious approaches fail)
Anthropic introduced three excellent tool use practices that I expect will be adopted widely.
- Tool search: Don’t pass the tool definitions to the model. Model can ask for a tool search when needed
- Programmatic tool calling: Instead of calling a tool, it’ll return a Python program to execute that will call the tools! This is a huge win
- Tool use examples: Lets you specific examples of tool calls to guide th model better
- The Hacker News thread flags that CLIs solve these - but CLI updates are hard, while APIs auto-update.
With AI, some skills that beome more valuable are (and will soon be in short supply, hence need to be taught) are: #
- Problem formulation (“What question should we actually ask?”)
  - Traits: Curiosity (absolutely), systems thinking, comfort with ambiguity, metacognition (thinking about your thinking)
  - Practice reframing exercises (“What are 5 other ways to frame this?”), study great questions in your field, work backward from outcomes, learn adjacent domains. The “5 Whys” technique helps. Also: deliberately pause before diving into solutions—force yourself to spend time in the question space.
- Taste and judgment (“Is this response appropriate?”)
  - Traits: Pattern recognition from experience, cultural literacy, empathy, contextual awareness, aesthetic sense
  - How to strengthen: Immerse yourself in excellent examples, study spectacular failures (they’re more instructive!), get feedback on your calls, practice explaining why you made a judgment. Build a “swipe file” of great/terrible examples. The key is volume—you need lots of reps.
- Quality assessment (“Is this AI output correct?”)
  - Traits: Healthy skepticism, attention to detail, domain knowledge, logical reasoning, understanding of edge cases
  - How to strengthen: Study common AI failure modes, build verification checklists, practice the “does this make sense?” test, learn what “good” looks like in your domain, cross-reference claims. Develop your “bullshit detector” by analyzing why wrong answers feel wrong.
- Creative synthesis (“How do these ideas connect?”)
  - Traits: Associative thinking, wide knowledge base, playfulness, comfort with non-obvious connections, intellectual courage
  - How to strengthen: Consume diverse inputs outside your field, practice analogical thinking (“X is like Y because…”), use visual thinking tools like concept maps, study how innovations happen in other domains, give yourself permission to make weird connections. Read broadly—fiction, history, science.
- Domain expertise (“Does this solution work in reality?”)
  - Traits: Deep curiosity, persistence, willingness to get hands dirty, learning from failure, long-term commitment
  - How to strengthen: Deliberate practice on real problems, seek mentorship, study edge cases and failure modes, build things (don’t just read about them), learn your field’s history. The “10,000 hours” thing is real, but it’s quality hours that matter.
- Meta pattern:
  - Reflection loops: doing something, then analyzing why it worked/didn’t.
  - Exposure to excellence: you can’t develop taste without seeing great work.
Some more new CLI tools I installed:
- trash-cli: Alias rm to move files to trash instead of deleting permanently.
After a week of seeing ligatures in Fira Code, all other fonts look ugly. My favorite ligatures: !== ==> =» <–> (and every possible arrow) >= ||> ||- |- …
The first name, alphabetically (at least among Straive employees) is “Aabida” and the last is “Zyrene”. Something I would never have discovered working in a smaller company.
chokidar-cli is an easy way to run commands when files change, e.g. npx -y chokidar-cli '**/*.js' -c 'npm run build'
npx -y mapscii shows a map on the terminal. Not too useful, not maintained, but very interesting.
termsvg converts asciinema .cast files to animated SVG suitable for embedding in GitHub (e.g. via mise x github:MrMarble/termsvg -- termsvg export file.cast --minify). The animated SVG is ~10X larger than the .cast file. The GZipped size is fine but saving it as .svgz is not recognized by GitHub. In contrast, agg, the official asciinema-to-GIF converter, creates .GIF files that are only 5X larger. The most efficient seems to be embedding via asciinema.org

usql queries MySQL, Postgres, SQLite, MSSQL, Oracle, etc via a single interface. For example, usql 'mysql://rfamro:@mysql-rfam-public.ebi.ac.uk:4497/Rfam' -c "SELECT * FROM clan limit 3;". But DuckDB is more versatile, IMHO.

INSTALL mysql; LOAD mysql;
ATTACH 'host=mysql-rfam-public.ebi.ac.uk port=4497 user=rfamro database=Rfam' AS rfam (TYPE mysql);
SELECT * from rfam.Rfam.clan LIMIT 3;
SELECT * FROM 'file.xlsx' LIMIT 3;
SELECT * FROM 'file.csv' LIMIT 3;

Autistic and allistic people just have different communication styles. Autistic people have no trouble understanding other autists. They just happen to be in a minority which makes it seem like they have a social deficit. Conflict between Neurotypes
1 second = 10 tokens for OpenAI Realtime APIs. 1 second = 25 tokens for Gemini Live API
- 39 cents / hour on GPT Realtime Mini = 36 cents audio input + 3 cents text output
- 139 cents / hour on GPT Realtime = 115 cents audio input + 15 cents text output
- 30 cents / hour on Gemini 2.5 Flash Native Audio (Live API) = 27 cents audio input + 3 cents text output
Here are some AI experiments I’m planning to try with our marketing team:
- Video Generation: Create marketing videos from text scripts in minutes
- Poster Generation: AI designs high-conversion posters from brief text inputs - notably Nano Banana Pro
- Synthetic Persona A/B Testing: LLM agents simulate 100K+ user behaviors to test designs before real users
- LLM-Powered A/B Automation: AgentA/B system runs experiments with AI-simulated traffic
- Vibe Coding Landing Pages: Marketers build production-ready pages in hours vs weeks
- On-demand Landing Pages: Generate pages for automated campaigns/products without human intervention
- Brand Voice Cloning at Scale: Train on company content to ensure consistency across 1000s of pieces
- Persona-Driven Content Synthesis: Use 1B+ personas to generate diverse content perspectives
- Competitive Intelligence Briefing: Real-time monitoring across millions of data points + data storytelling
- Marketing Analytics with LLMs: AI agents analyze complex datasets for insights
- Brand Compliance Checks: Ensure all content meets brand guidelines automatically
- Autonomous Blog Squads: AI agents identify trending topics / internal content, create data stories ready for review
New skill unlocked: creating tutorials from talk proposals. I asked Claude to Write a Malcolm Gladwell article based on this talk description to teach me the topic and passed it this talk proposal: Your Causal Parrot might be lying to you. The story it wrote is very engaging and informative!
- LLMs “understand” causality because of training, but lack a world model to extrapolate to new situations.
- Giving them tools to reason (e.g. causal models, sub-agents to explore root causes) will help.
A cool Gemini 3 Pro hack: convert satellite imagery into stylized maps! Bilawal Sidhu
Running sub-agents in tmux helps avoid timeout cancellation, and hence allowing resuming Peter Steinberger

Things I Learned - 26 Oct 2025

Sun, 26 Oct 2025 00:00:00 +0000

This week, I learned:

Before founding a place to do good, work in a place that does good and learn. Ben Werdmuller
What should we teach when vibe coding becomes good enough for non-coders? Ethan Mollick
- Problem decomposition
- Clear communication & spec writing
- Core technical foundations: file systems, access control, networking, APIs, version control, data structures, databases, deployment
- Software development skills: Debugging, Testing, Refactoring, Design patterns, UI/UX
- Project management: requirements, prioritization, scoping, …
Codex CLI tips:
- codex --add-dir $DIR lets you write into $DIR
- codex --full-auto is the equivalent of codex --sandbox workspace-write --ask-for-approval on-request
Terse code is not necessarily easier or harder for LLMs to write. It’s about how unusual (or not aligned with training data) the code is. Gabi Teoduru
How are people using browser agents like Comet / Atlas? Simon Willison
- Most popular: YouTube video summaries with timestamps
- Most useful: Form filling: Government forms, data entry, repetitive bureaucratic tasks
  - Foreign language navigation: Applying for pension in Korea, navigating sites in other languages
  - Time reporting auto-completion
  - Insurance claims: Reading policy documents and drafting appeals (successfully got claim reimbursed in India)
  - Compliance training click throughs
- Next most useful: Shopping / planning
  - Energy provider comparison - Comet checked current plan vs competitors on Check24, calculated exact annual savings per provider
  - Financial tracking: Finding Amazon orders, tracking Airbnb spending with refund calculations, analyzing bank transactions
  - Trip planning: Mapping 50-100 places on Google Maps automatically
- Interesting: Airport shuttle discovery - Found shuttle that user missed in manual searching
HubFS mounts GitHub repos on the file system. Every file system action directly works on GitHub via a REST API. Useful for some scenarios but less useful for note-taking than something like GitDoc which offers a delayed sync.
Ernest Ryu solved an open problem in convex optimization using ChatGPT. Quotes:
- ChatGPT is now at the level of solving some math research questions, but you do need an expert guiding it.
- ChatGPT was really effective at accelerating my progress. This work took about 12 hours, spread over 3 days. In hindsight, the proof is really simple.
- But I iterated through so many other strategies that didn’t pan out, and ChatGPT crucially helped to quickly explore and eliminate those dead-end approaches. Also, the key successful steps were suggested by ChatGPT.
- ChatGPT did not produce the proof in a single prompt. The process was highly interactive. It generated many arguments, roughly 80% of which were incorrect.
- Yet some were genuinely novel to me. Whenever I recognized a novel idea, whether correct or only partially so, I distilled the key insight and prompted ChatGPT to develop it further.
- My contribution:
  - Filtering out incorrect arguments and accumulating a set of correct facts.
  - Identifying promising new lines of reasoning and guiding ChatGPT to explore them further
  - Recognizing when a strategy had been fully explored and deciding when to move on.
- ChatGPT’s contribution:
  - Producing the final proof argument.
  - Significantly accelerating my (or our) exploration of the many dead-end arguments, rapidly ruling out approaches that did not work.
Comparing the GPT 4.1 and 5 models at all different of reasoning, I’ve switched my default from GPT 4.1 mini to GPT 5 mini (medium). Far smarter for a slightly higher cost. Artificial Analysis
python -m pdb -c continue script.py or uv run -m pdb -c continue script.py runs a script and drops into pdb on unhandled exceptions (post-mortem). ChatGPT
Technology removes constraints. We then do what we really value. Claude
- When writing became digitized, we stopped cared about spelling/handwriting for its own sake. Spelling bees and handwriting classes declined. “ur” is acceptable.
- When fitness tracking became easy, many just track, few exercise more. Few people value exercise
- When GPS became ubiquitous, we stopped learning geography. Most value arriving, not knowing
- When photography became unlimited, most captured moments. Few perfected shots
I had Codex scrape my ~2,000 pending invites on LinkedIn and asked ChatGPT to analyze it. Here are learnings: ChatGPT, private
- Power-law. 5% of inviters account for ~42% of all common connections. Top 10 people alone for ~20%.
- IITM student invites are high (~14%), but with 0-2 common connects, i.e. distant strangers.
- EdTech is tiny in count but has the highest common connections per person (outlier-sensitive but real).
- Among ≥20-commons, many hold VP/Head/Site-Lead titles in Data/AI or GenAI (not just recruiters).
- GenAI people are 7-8% and steady across months. Not a useful signal to prioritize.
- Premium ~ Senior. Premium accounts show ~40% senior titles vs ~29% for non-premium.
- Finance invites have higher seniority rate and more common connects than healthcare.
- Followers have higher common connections (~6 vs ~4).
⭐ Memory can be code. Agent memory is anything it choose to persist. Agents can write code on the fly to automate tasks, save them, and serve the code on the next request, potentially modifying the code as required. This is like the conscious mind saving a habit for the subconscious to execute fast.
Finally: Microsoft Office has an agent mode that lets you talk to it and do stuff. The Verge