AI prices and spending
Epoch AI estimates benchmark scores got about 13x cheaper a year. Anthropic says Claude Opus 5 comes close to Fable 5 at half the price, and it priced Opus 5.5 at 20% less than Opus 5. An Army email says its unlimited AI token pool ran dry by mid-June.
What organizations and people pay to use AI: model and plan prices, usage limits, usage-based billing, spend caps and company AI budgets.
October 2026
The facts
DHH says OpenAI's Sol finished the Agents on Rails test tasks less than 5 percentage points behind Opus, at $0.80 versus $10.53.1
Artificial Analysis says Google's Gemini 4 Argon ranks equal to GPT-6 Astra on its index, at 60% of the cost per task while its prices are cut in half until November 30.2
Coinbase CEO Brian Armstrong says Coinbase fine-tuned an open model, Qwen3.5-9B, on its own data, and it beat frontier models at fraud prevention, at much lower cost.3
Accenture has told many employees it will now spend only $29 a month per person on Claude. Some had been spending as much as $1,000 a month.4
OpenClaw creator Peter Steinberger, now at OpenAI, estimates he spent around $10 million on AI use (tokens) last month. He says not all of it is useful.5
Browser Use says DeepSeek V4.1 Flash 1017 did as well as Opus 5.5 on long tasks in a web browser, and cost $0.22 a task versus $2.35.6
The stories
- Kimi K3.1 Pro is priced at $2/M input and $10/M output, down from K3's $3/$15, listed 50% below Opus 5.5.
- SemiAnalysis claims Anthropic is A/B testing Claude usage limits, lowering monthly API-equivalent value about 20% while raising 5-hour limits.
- OpenAI adds 12-hour API prompt caching, mixable with the 30-minute default within a request, at extra cost for longer retention.
- OpenAI identifies about 40,000 accounts whose applied usage limits did not match their plans, fixes it and promises credits.
- A user reports Claude Opus 5.5 finishing Baldur's Gate 3 in 39 hours for $712, with 99 sessions and 20,809 tool calls.
- OpenClaw creator Peter Steinberger of OpenAI says at Cloudflare Connect 2026 he spent around $10 million in tokens last month, adding not all of that is useful.
- Artificial Analysis lists DeepSeek V4.1 Flash 1017 as a 552B-total, 16B-active MoE with 1M-token context and MIT license, pricing unchanged at $0.30/$1.20 per million tokens.
- Browser Use reports DeepSeek V4.1 Flash scores 75/100 on its long-horizon browser test, matching Opus 5.5, at $0.22 versus $2.35 per task.
- MiniMax launches M3.1 Flash at $0.15/$0.60 per million tokens, claiming 1/3 the cost and 3x the speed of M3; weights arrive Oct 23.
- Accenture tells employees they can spend only $29 a month on Claude, 404 Media reports.
- Anthropic prices Haiku 5.5 at $0.10/$0.50 for prompts up to 100K tokens, about 90% below Haiku 4.5, with 39.2% on Terminal-Bench 4.0 at max effort.
- Elorian launches Glimpse, a visual-thinking VLM, and claims it matches or beats Astra on 5 of 12 benchmarks at 17x lower token prices.
- Liquid d1-ultra scores 65 on Artificial Analysis's Decision Index versus 57 for Jev 1.14, at 92ms per task and $0.042 per M input tokens.
- Perplexity releases pplx-decider v1.2 open weights: a 27B model at 2 cents/M and a 4B Flash at 1 cent/M, with up to 5x faster single-pass multi-question inference.
- Together, Fireworks and Baseten offer Muse Spark Open at $1.25/$4.25 per 1M input/output tokens; Together runs a free endpoint until Oct 18.
- Anthropic makes Opus 5.5 fast mode GA with no waitlist: up to 2.5x speed at $8/$40 per 1M tokens.
- Krea releases Krea 3, claiming 2K edits in under 8s at $0.06 per image.
- Ultramarine 1 Flash, an open knowledge-work model, is claimed to match Fable 5 at 60x lower cost; the API is live and weights are due in two weeks.
- DHH reports GPT-6.1 Sol within 5 points of Opus on his task suite at $0.80 versus $10.53.
- Artificial Analysis puts Argon level with GPT-6 Astra on its Intelligence Index at about 60% of Astra's cost per task, with a 15% AA-Omniscience hallucination rate.
- ScaleDown launches Classify, Extract and Summarize endpoints at $0.05 per million input tokens and claims 200x speed and 467x cost advantages over frontier APIs.
- Artificial Analysis now publishes measured cache-hit rates and provider-specific cost per task; GLM-5.3 Max spans $0.92-$8.68 per task.
- Mark Zuckerberg says Meta is at the "frontier", that safety is each lab's job, and that superintelligence costs up to $200 a month.
- Wonder and Wonder Full drain Muse Code quotas abnormally fast on launch day; Meta fixes it the same day and resets quotas for all users.
- Artificial Analysis puts Wonder cost at $4.52 per index task; in its Coding Agents Index Wonder Full in Muse Code scores 68, tying Claude Code with Sonnet 5.5.
- River AI launches inference 20% cheaper than first-party DeepSeek V4.1 Flash and GLM-5.3 Flash, plus on-prem custom training.
- SubQ Snap launches with 384K context at $0.05 per million input tokens and free output.
- Cursor joins the X Subscription; paid plans share one usage pool with Grok Build, and third-party models in Cursor move to API pricing.
- LangChain says routing each task to the cheapest adequate model cut median Open SWE cost per task by 64%.
- Opus 5.5 fast mode rolls out; it bills against usage credits and is not included in subscriptions.
- Vals AI finds agent teams of GPT-6 Sol and Opus 5.5 cost 1.8-5.1x more on Vibe Code Bench; only Sol at medium effort improves significantly, by 7.3 points.
- Tinker cuts prices by up to 70%, prices long-context the same as short, and adds GLM-5.3-Flash and DeepSeek-v4.1-Flash.
- An a16z report finds only 4.5% of US consumers pay for gen AI apps.
- Anthropic releases Claude Haiku 5.5 with 1M context, 128K max output and $0.10/$0.50 per million token pricing; Anthropic says about 75% lower than Haiku 4.5.
- Anthropic halves Sonnet 5.5 cache reads to $0.10 per million tokens and estimates most agentic work gets about 20% cheaper.
- Anthropic adds monthly Claude Platform API credits to Max 5x ($100), Max 20x ($200) and Team (up to $500, pooled) plans.
- Cognition reports Claude Haiku 5.5 scores 58.4% on FrontierCode 1.1 in Devin, ahead of Sonnet 5 at roughly one-eighth the cost per task.
- Cursor claims Claude Haiku 5.5 is "10x less than Haiku 4.5" on shorter requests.
- SemiAnalysis finds Claude subscriptions give 5x+ more API-equivalent value than OpenAI plans, or 1.3-2.9x after adjusting for task cost.
- Cline previews Pareto 26.10, a model router claiming $0.24 versus $13.41 per task at equal DeepSWE score.
- Epoch estimates OpenAI researchers' coding-agent spend at API prices doubled roughly monthly, to about $600 a day for the median by mid-August.
- Vals finds Jev matches GPT-6 Astra's 97.5% on claim verification at about 1/500th the cost, but ranks last on LegalBench.
September 2026
Epoch AI estimates benchmark scores got about 13x cheaper a year. Anthropic priced Claude Opus 5.5 at 20% less than Opus 5.
35 stories in the AI newsletters I read.
The facts
Epoch AI estimates benchmark scores got about 13x cheaper a year.1
Anthropic priced Claude Opus 5.5 at 20% less than Opus 5.2
All 35 stories
- OpenAI releases GPT-6.1 Sol, which nearly matches GPT-6 Astra at one-fifth the price.
- OpenAI adds a 30-minute prompt cache-hit guarantee, cache pre-warming, a diagnostics tool and 25% cheaper cache reads.
- OpenAI re-tiers ChatGPT plans and adds Pro 500 at 25x, roughly halving old Pro 200 value.
- Artificial Analysis measures GPT-6.1 Sol at $0.72 per task at max effort, versus $3.26 for Astra.
- Google prices Gemini 4 Argon at $4/$20 per million tokens, halved to $2/$10 with no end date announced.
- Open-weight models reach 56% of Vercel's traffic as firms chase cheaper tokens.
- The cost of 75% on ARC-AGI-1 falls 99.95% in 19 months, from $26 per task to a penny.
- Anthropic releases Claude Sonnet 5.5, 30%+ faster and up to 30% cheaper than Sonnet 5, scoring 70.6% on Terminal-Bench 4.0.
- Claude Sonnet 5.5 ranks second on the Artificial Analysis index but costs 50% more per task than Sonnet 5.
- Anthropic resumes billing for safeguard-blocked requests, with a false-positive rate under 0.1%.
- A Jev-as-a-Judge paper reports $0.044 per 1K judgments at 152ms, about 277x cheaper than GPT-6.
- Ramp matches GPT-5.6 Luna reranking accuracy with Jev at 10x lower tail latency and 3x lower cost.
- UkisAI releases Swift models that cut thinking tokens by up to 63.4% with little score change.
- Anthropic ships Claude Opus 5.5, cuts list price 20% to $4/$20, and says it runs 40% cheaper than Opus 5 at default effort.
- Anthropic raises 5-hour session limits 20% and gives Pro, Max and Team users a banked rate-limit reset.
- OpenAI offers up to 90% discount on cached input reads and adds a Prompt Caching Dashboard and diagnostics API.
- Epoch AI finds the price of a given AI capability falls 13x a year.
- OpenAI releases GPT-6 Sol and Luna at about half their predecessors' API prices and says Sol beats Opus 5 at 9% of the cost.
- xAI releases Grok 4.7, twice as fast at half the cost, though it burns triple Astra tokens.
- Legal AI startup Harvey, at -50% margins as token use rose 20x, turns positive on Kimi K3 as startups move to open weights.
- Braintrust offers Jev as an eval model at about 400x lower scoring cost.
- Roboflow rates GPT-6 Astra its strongest vision model and adds it to Auto Annotate, at double the per-image cost on high effort.
- Google updates Gemini managed agents with an Antigravity-based harness, claiming up to 30% lower cost and 22% higher cache hits.
- Steve Yegge shuts down Gas Town, saying thousands of dollars a month in coding-agent subscriptions built only that project.
- Databricks rolls GPT-6 Astra out to about 3,500 engineers after a 200-user pilot; coding spend rises about 60%.
- Union Alpha becomes free in Cline with 256k context, claiming Astra-class coding at about 18x lower cost.
- OpenAI cuts desktop voice pricing about 60%, with usage up 2.4x, and adds ChatGPT gift cards.
- Cognition's SWE-2 trails Fable by one point at 64% less cost.
- DeepSeek prices V4.1-Flash at $0.30 per 1M input and $1.20 per 1M output tokens, with 50% off-peak.
- OpenAI pauses new $200 Pro signups because of Astra capacity; existing users, the API and other plans are unaffected.
- LlamaIndex launches LlamaParse connectors for Claude and ChatGPT as a cheaper option for bulk document extraction.
- Perplexity integrates Fable 5.1 into Perplexity Computer as an orchestrator, citing a 21% higher WANDR score at 37% lower cost than Fable 5.
- Anthropic releases Claude Fable 5.1 and Mythos 5.1, one model at two safeguard tiers, with a 75% cache-read price cut for Fable.
- Sonar Vortex gives agents a semantic code graph and cuts task cost 5-36% versus text-search workflows.
- Anthropic Max plan weekly caps draw the most viral customer backlash of the day.
August 2026
Google offered Gemini 3.7 Flash at an introductory price of $0.75 per million input tokens.
21 stories in the AI newsletters I read.
The facts
Google offered Gemini 3.7 Flash at an introductory $0.75 per million input tokens.1
All 21 stories
- Salesforce puts its whole CRM inside Claude with 37 sales skills, billed by consumption.
- Perplexity launches Portable Computer, running a full agent stack on a DGX Spark with no per-token charge.
- OpenAI ships GPT-5.6 in Kiro with about 82% lower cost per task.
- GPT-5.6 Sol Max scores 72.7% on DeepSWE v1.1 at $6.47 per task, against Fable 5 Max's $21.63.
- AT&T routes 40% of employee AI usage to open models, cutting coding costs 56%.
- Ramp Router offers GPT-5.6 Sol at 50% off for a limited time, amplified by GitHub Copilot and VS Code.
- OpenAI cuts GPT-5.6 Sol pricing by over 20% for three months.
- LangChain ships LangSmith Tuned Evaluators, claiming better-than-frontier results at 82% lower cost.
- Glean claims $0.45 per task versus $1.84 for Claude Code and Cowork.
- Google ships Gemini 3.7 Flash three weeks after 3.6, at half the price, to power its Spark agent.
- DeepSeek raises API prices from Aug 16 and adds 2x peak billing; V4-Pro output goes from $0.87 to $1.98.
- Data shows the median company spends $12 per employee per month on AI, while the top 1% spend $7,500.
- Anthropic makes Claude Sonnet 5 pricing permanent at $2/M input and $10/M output.
- Top corporate AI adopters burn 8.3 times the median firm's tokens.
- Gartner forecasts AI-optimized cloud spending up 96% in 2026, to over $42 billion, with inference at $23 billion exceeding training at $19 billion.
- Databricks cuts internal AI coding spend up to 90% in some scenarios while usage grows.
- Microsoft caps engineers' AI spend and says tokenmaxxing is not the goal.
- Not Diamond Code launches a per-step model router for coding agents, claiming 20-65% cost reduction.
- DeepSeek V4 Flash is named the cheapest model above 60 on the Vals Index, 35x cheaper than the next best.
- DeepSeek V4-Flash runs at $0.14 and $0.28 per million tokens and scores 50 on the Intelligence Index.
- OpenAI's CFO touts an 80% price cut, a billion users, and agents writing 99.8% of output tokens.
July 2026
Anthropic says Claude Opus 5 comes close to Fable 5 at half the price. An Army email says its unlimited AI token pool ran dry by mid-June.
29 stories in the AI newsletters I read.
The facts
Anthropic says Claude Opus 5 comes close to Fable 5 at half the price.1
An Army email says its unlimited AI token pool ran dry by mid-June.2
All 29 stories
- OpenAI cuts GPT-5.6 Luna prices 80% and Terra 20%, and adds a Fast mode that runs 2.5x quicker for twice the price.
- Buyers go a la carte on cheaper AI models, some Chinese.
- Cline signs the Open Weights letter and makes GLM 5.2 free in Cline.
- OpenAI's GPT Transcribe scores 3.31% word error rate, priced 25% lower at $4.50 per 1,000 minutes.
- Nous Research adds Opus 5 to its Nous Portal with a 20% discount across all models.
- Anthropic releases Claude Opus 5, Fable-class intelligence at half the price.
- Public benchmark data shows some Chinese models within 3% to 18% of Western results at five to 12 times lower cost.
- Cursor launches Cursor Router, claiming frontier-quality results at 60% lower cost than routing everything to Opus 4.8.
- OpenAI rolls out hard spend limits to all API accounts.
- Meta's AAI Labs is reportedly building its own model router to send tasks to cheaper models.
- SambaNova ships prompt caching in SambaCloud, claiming 90% cheaper cached tokens and up to 91% lower time to first token.
- Microsoft is reportedly moving Kimi K3 onto Azure to cut up to $600 million in inference costs.
- The US Army uses a year of unlimited AI tokens in six weeks.
- Ramp launches Ramp Router, an OpenAI-compatible endpoint that routes across GPT, Claude, Gemini, Grok, Qwen, DeepSeek, Kimi and GLM.
- Alibaba upgrades its Qwen Token Plan to bundle the whole model family at 40% off, with Claude Code and Cursor support.
- Artificial Analysis scores Kimi K3 a 13-point intelligence jump at $0.94 per index task, half the cost of Opus.
- Muse Spark 1.1 beats GPT-5.6 Sol on 525 clinician tasks at a seventh of the price.
- Elon Musk orders Tesla staff to move onto Grok, citing token cost.
- OpenAI, Meta and SpaceXAI ship models led by cost, with GPT-5.6 using fewer tokens and Grok 4.5 twice as efficient.
- OpenAI resets usage limits twice after the GPT-5.6 launch and promises to restore familiar ChatGPT sidebar patterns.
- OpenAI claims GPT-5.6 Luna at lowest effort beats GPT-5.5 at highest effort on health tasks at 25x lower cost.
- Perplexity adds Grok 4.5 as an orchestrator in Computer at roughly half the cost of Opus 4.8.
- Anthropic moves Fable onto premium credits at $10 in and $50 out per million tokens.
- A study finds changing only the orchestration layer cuts cost per task 41% and tokens 38% at equal quality.
- Hugging Face's Reachy Mini team swaps GPT-realtime, quoted at $45k a month, for an open voice stack at $0.25 an hour.
- GPT-5.6-sol ties Claude Fable 5 for #1 in frontend coding at half the price.
- Meta launches Muse Spark 1.1, its first paid model, at a quarter of rivals' prices.
- Anthropic resets all rate limits in the middle of the GPT-5.6 launch.
- Grok 4.5 debuts at #4 on GDPval-AA at a tenth of the cost of higher-ranked models.
June 2026
California agencies can get Claude at a 50% discount per seat.
25 stories in the AI newsletters I read.
The facts
California agencies can get Claude at a 50% discount per seat.1
All 25 stories
- Cognition's Devin Fusion cuts coding task costs 35% by routing subtasks to cheaper models while keeping Fable-level quality.
- LlamaIndex ships a Retrieval Harness and a Trace Judge model that detects agent errors at 1/100th the cost of closed-model judges.
- Cline launches a $9.99 per month pass covering GLM 5.2, DeepSeek, Kimi, MiniMax and Qwen.
- Governor Newsom gives every California agency Claude at a 50% discount plus free workforce training.
- UBS data shows 60% of companies with AI budgets are shifting to cheaper or open models.
- Coinbase nearly halves AI spend while usage rises, using cheaper defaults, routing and caching that lifted hit rate from 5% to 60%.
- GLM-5.2 scores 22.8% on ARC-AGI-2, the best open-source result so far, at about a third of Opus 4.8's cost.
- Gartner forecasts AI coding costs will pass the average developer salary by 2028 as token use surges.
- Cline finds GLM-5.2 fixes a real repo bug at half the cost of Opus 4.8.
- LangChain pushes LangSmith Engine, with a fine-tuned trace judge at 10-100x lower cost than frontier models.
- Microsoft adds a cheap DeepSeek V4 tier to Copilot Cowork.
- OpenRouter launches Fusion; a budget trio of models beats GPT-5.5 and Opus 4.8 at half the cost.
- Meta tells staff it will cap AI token usage and steer engineers to its own MetaCode.
- OpenAI considers drastic price cuts in anticipation of a war for users with Anthropic.
- LangChain's LangSmith LLM Gateway adds spend limits, PII and secrets detection, trace continuity and audit logging.
- Google cuts Google AI Plus pricing from $7.99 to $4.99 a month and adds agentic chat to NotebookLM.
- Cloudflare adds AI Gateway spend limits with fallback to cheaper models when caps are reached.
- Anthropic doubles Claude Cowork usage limits for one month.
- Enterprises route hard tasks to frontier models and easy ones to cheaper rivals, pressuring OpenAI and Anthropic.
- Harvey finds a GLM 5.1 plus Opus 4.7 legal agent beats pure Opus, 18% vs 14%, at $368 vs $954.
- Uber caps engineers at $1,500 a month per AI coding tool after burning a year's budget in four months.
- Anthropic launches Claude Sonnet 5, aimed at agentic apps, with capabilities close to Opus 4.8 at a lower price.
- GitHub Copilot users are dismayed by usage-based billing; O'Reilly expects it to raise developers' payments significantly.
- Developers report Claude Fable is much better than Claude 4.8 but burns tokens fast and has strict guardrails.
- Databricks launches Unity AI Gateway, tools for organizations to manage AI costs.
May 2026
The stories
- Anthropic's Opus 4.8 Fast mode runs 2.5x faster at 2x the normal price, down from 6x for Opus 4.7 Fast.
- Xiaomi MiMo cuts its cache-hit price, with about 80% lower caching cost.
- Meta tests paid AI features for the first time, from $7.99 a month.
- Artificial Analysis measures Gemini 3.5 Flash at up to about 280 output tokens/s, at about 5x the cost of Gemini 3 Flash.
- Artificial Analysis puts DeepSeek V4 Pro pricing at $0.435/M input and $0.87/M output, 19x cheaper to benchmark than Opus 4.7.
- DeepSeek makes the 75% DeepSeek-V4-Pro discount permanent.
- Artificial Analysis reports Cursor Composer 2.5 is 3-18x cheaper than Opus 4.7 on coding agent benchmarks.
- OpenRouter finds GPT-5.5 raised its token price but cut tokens per typical conversation; Claude Opus 4.7 needs more tokens at the same price.
- Lanai launches TokenTuner, which finds where companies can use lower-cost models and rates teams on token use.
April 2026
March 2026
November 2025
The board, once a month, by email.
It comes on the first Monday of the month.
By subscribing, you agree to get the board by email.See the Privacy Notice.