One Balance Sheet, Two Very Different AI Bets
Microsoft's latest earnings quietly disclosed the clearest verdict yet on its two biggest AI investments, and they are heading in opposite directions.
When Microsoft reported fourth-quarter earnings for its fiscal 2026 year (ended June 30), it tucked in an unusual disclosure: how its investments in the two biggest, competing AI labs are actually performing on paper. For the quarter, Microsoft recorded its stake in Anthropic as a $3.2 billion gain, adding 33 cents to diluted earnings per share. Microsoft invested $5 billion in Anthropic in November 2025 as part of a deal under which Anthropic also committed to buy $30 billion of Azure services.
Its OpenAI stake told a different story. Microsoft marked that investment down about $600 million for the quarter, cutting diluted EPS by roughly 7 cents. Microsoft owns about 27% of OpenAI and, unlike Anthropic, updates that valuation every quarter. The dollar swing itself is a rounding error against a quarter in which Microsoft reported $90 billion in revenue and $35.8 billion in net income, but the disclosure itself is the story: Microsoft chose to highlight a one-quarter Anthropic gain that came close to matching its entire full-year OpenAI gain.
$3.2BAnthropic Gain, Q4 Alone
-$600MOpenAI Markdown, Q4
$5.0BOpenAI Gain, Full Year
Microsoft doesn't routinely revalue its Anthropic stake and rarely calls out these numbers at all. Choosing to disclose a $3.2B quarterly gain, one quarter away from matching a full year of OpenAI gains, is itself a signal about which bet Microsoft wants investors watching.
The read for operators: this isn't a story about picking winners between frontier labs. It's a reminder that the biggest, best-capitalized company backing your AI vendor's ecosystem can flip its own internal read on that vendor from quarter to quarter. If your automation roadmap leans on any single AI relationship for the long term, that relationship's own backers are re-pricing it in real time, and you should know which way.
Prices Are Falling. Your AI Bill Isn't.
Token prices are in genuine free-fall as cheaper models take share. Enterprise AI spending is still climbing anyway. Both things are true at once, and the reason matters for anyone budgeting AI this year.
Enterprise AI providers are facing what amounts to their first real pricing reckoning. Companies are increasingly routing routine work to cheaper models, many of them Chinese, rather than defaulting to the most expensive frontier option for every task. Chinese suppliers, including DeepSeek, Qwen, GLM, Kimi, and MiniMax, have seen their share of US enterprise token usage climb as high as 46% at points this year, per OpenRouter usage tracking, up from under 5% in early 2025, and it has held above 30% every week since early February. That shift is squeezing margins at costlier U.S. labs and pushing some vendors, including Cursor, to adjust pricing downward. In one illustrative example cited by Forbes, swapping to a cheaper model for a routine coding task (building a basic web browser) cut the cost by roughly 87%.
None of that is translating into lower bills, though. The reason is consumption, not price. Gartner's analysis this year found agentic workflows require five to 30 times more tokens per task than a simple chatbot query, since an agent plans, calls tools, and iterates rather than answering once. That volume increase is outrunning the price decline for a lot of operators, even as the sticker price per token keeps dropping.
↓ What's Actually Falling
- Per-token list prices across most vendors
- Cost of routing routine tasks to cheaper models
- Cursor and other resellers' own list prices
↑ What's Actually Rising
- Total tokens consumed per completed task (agentic vs. chat)
- Total monthly AI spend at most enterprises
- The gap between what was budgeted and what was billed
six50's Token Cost Governance Checklist for SMB Operators
1
Track total tokens consumed per completed workflow, not the per-token price your vendor advertises. Price is falling, consumption is what's actually driving your bill.
2
Route routine, high-volume tasks to the cheapest model that clears the quality bar. Save premium models for judgment calls that actually need them.
3
Before approving any new agentic workflow, model its per-task token cost against the equivalent chat-based version. A 5-30x consumption jump is the default, not the edge case.
4
Revisit model routing every quarter. A vendor's pricing tier from Q1 can already be stale, in either direction, by Q3.
The read for operators: this is exactly the blind spot our token and model routing cost governance work is built to catch. A falling price list can mask a rising bill just as easily as a rising price list can, and most SMB finance teams are only watching one side of that ledger.
An Unreleased Claude Model Just Cracked Real Cryptography
Anthropic says its research-only Claude Mythos Preview found a genuine weakness in a NIST post-quantum candidate. No production system is affected, but the capability signal is real.
Anthropic disclosed that Claude Mythos Preview, an unreleased research model, derived an improved key-recovery attack against HAWK-256, one of the lattice-based signature schemes competing for post-quantum standardization, by finding a previously unexploited mathematical symmetry in its underlying lattice. The attack drops the cost of recovering HAWK's smallest key from roughly 2^64 operations to about 2^38, a very large practical reduction, though still an exponential-time attack, not a full break, and not applicable to other NIST candidates or to lattice cryptography broadly. Separately, the same model developed a new technique, which Anthropic calls the "Möbius Bridge," that produces a 200-to-800-fold speedup on attacks against a deliberately weakened, seven-round version of AES-128 (full AES-128 uses ten rounds and is unaffected).
2^38Ops to Break HAWK Key (was 2^64)
200-800xSpeedup, Weakened 7-Round AES
~$100KAPI Cost, Per Result
Per Anthropic's own account, the HAWK attack took about 60 hours of work to find, develop, and verify end-to-end. The AES speedup took roughly a week for Mythos to conceive, though it then took Anthropic's own researchers, who are not cryptography specialists, nearly a month of additional work to fully verify the result was correct. Anthropic put the API cost at roughly $100,000 for each of the two results. HAWK has not been standardized or deployed anywhere in production, and Anthropic states plainly that neither result requires any change to production software today.
The read for operators: nothing here is an active threat to any system you run. What's worth tracking is the capability curve, not the specific result: a research model did in about 60 hours, for roughly $100K in compute, work that survived two years of expert human review before Mythos found the gap. That is the same kind of quiet capability shift we watch for when we scope which parts of a client's stack are worth automating now versus in a year.
The six50 POV
What we'd tell a $2M-$50M operator to do with today's news
01 — VENDOR CONCENTRATION RISK
A single quarter swung Microsoft's own internal read on OpenAI from gain to markdown. If your AI automation roadmap leans on one vendor relationship for anything mission-critical, that vendor's own biggest backer is re-pricing the bet quarterly. We build vendor diversification into every First 90 Days Diagnostic for exactly this reason.
02 — TRACK CONSUMPTION, NOT LIST PRICE
Falling token prices are not the same thing as a falling AI bill. Agentic workflows can burn 5-30x more tokens than a chat query, which is exactly why we measure clients on cost per completed workflow, not the sticker price of the model underneath it, in our token and model routing cost governance work.
03 — THE MARKET IS SCALING FASTER THAN MOST BUDGETS ASSUME
OpenAI crossing $1B in monthly revenue is a market-maturity signal, not just a headline. When frontier labs scale this fast, the compounding spend shows up in every vendor downstream. This is why we now treat an AI-cost line item as a standard, not optional, part of a fractional CFO close-process engagement.
04 — DON'T ACT ON THE PACING DEBATE YET
OpenAI and Anthropic backing the pacing letter as companies is a bigger signal than the employee letter alone, but it's still a policy story before an SMB compliance one. Unless you're already in a regulated vertical, watch this over the next quarter rather than changing anything this week.