The lease question is small. The reads of it are not. Public reporting on Anthropic's sublease from SpaceX has produced at least three durations in circulation, and none of them are sourced to a filed document. The company shipped Claude 4.7 Opus on 2026-04-15 — 1M context tier, $15 and $75 per million tokens in and out — a release cadence that implies a headcount and a compute footprint someone has to seat. The lease term is the proxy for that footprint. Below: three composite reads of what the term might be, and what each one implies for the eighteen months after April.
The reads are not equally weighted in the trade press. The short-term read leans on whisper-network sourcing from commercial brokers in the South of Market submarket. The decade read leans on what is structurally normal for Class-A tech sublets in San Francisco when the sublessor is itself a tenant on a long primary lease. The optionality read leans on what large AI labs have actually signed in the last eighteen months when their own valuation curve is still vertical. We are not picking one. We are walking through what each one would mean if it turned out to be the contract that actually exists.
Scenario 1: The Short-Term Reader (24-Month Sublease, Bridging to a Larger Lease)
Imagine the lease term reported by the most aggressive of the three sources turns out to be the correct one: a 24-month sublease, expiring sometime in the second quarter of 2028, with no embedded renewal option. Picture a real estate desk at the company that approached the SpaceX space the way most growth-stage AI labs approach any sublet — as a bridge, not a destination.
The math on this read is the most legible. A 24-month term is what a tenant signs when they expect to outgrow the footprint before the term ends. It is also what a sublessor offers when they want optionality on their own primary lease, typically because they are themselves uncertain about headcount trajectory. In SOMA Class-A submarket, sub-three-year subleases priced in the $70 to $95 per square foot full-service range in late 2025 according to broker market reports — well below direct-deal pricing of $100 to $130 in the same buildings.
For a company shipping a model at the price tier of Claude 4.7 Opus — $15 per million input tokens, $75 per million output tokens, 1M context window — the implied operational scale is not 24 months of static headcount. SWE-bench Verified at 82.4 percent as of release, GPQA Diamond at 91.2 percent, HumanEval at 94.0. Those numbers are not produced by a research team that fits in a single floor plate. They are produced by a multi-cluster compute org with embedded product, safety, and platform engineering, and that org grows. The bridge read says: Anthropic took the SpaceX space because it was available, not because it was the right shape, and the 24 months buys them time to negotiate a 250,000-plus square foot direct deal somewhere else.
The read implies three downstream outcomes for the next 18 months. First: continued broker chatter through 2027 about which large floor plate Anthropic is touring — this is the kind of search activity that leaks. Second: hiring slows or stays flat in the city itself, with new headcount steered toward existing satellite offices. Third: the company's next funding round prices in real estate as a near-term capex line item rather than a deferred one, because a 250k-plus direct lease in the city carries a TI allowance negotiation, an LOI cycle, and a build-out timeline measured in eighteen-plus months from term sheet to occupancy.
If you are watching for this read to validate, the tell will be a press leak in 2026 Q4 or 2027 Q1 about a separate large SF lease — not a sublease — at a different address. The 24-month interpretation only survives if that second signing happens.
Scenario 2: The Decade Reader (10-Year Term, Standard Class-A Commercial Lock-In)
Now picture the opposite end. The lease is what San Francisco Class-A leases overwhelmingly are when the tenant is investment-grade and the building is institutional-owned: a 10-year primary term, two five-year renewal options, fixed escalators at three percent per annum, free-rent abatement front-loaded into the first twelve months, and a tenant improvement allowance in the $90 to $130 per square foot range.
In this read, Anthropic is not a sub-tenant in the loose growth-stage sense. The company is a sub-tenant in the structural sense — SpaceX holds the primary lease on the building, and Anthropic took an assignment-equivalent sublease that mirrors the primary terms. This is the read that brokers default to when the sublessor is itself a sophisticated tenant that wants to monetize unused floors without exiting the building. The trade press has not picked this read up loudly, partly because it is the least narratively interesting one. A ten-year lease is not news. It is the base rate.
The implication for the eighteen months after April is the inverse of Scenario 1. If the lease term is ten years, the company has effectively pre-committed to a headcount range — not exact, but bounded. The space tells you how many seats the company expects to fill on its first day in the building, with growth assumed to fold into the building as floors clear of other sub-tenants. The signal travels through hiring posture: a ten-year lease justifies aggressive in-region hiring, because the seat cost is amortized over a horizon that absorbs even pessimistic productivity assumptions.
It also changes how the company budgets compute relative to people. Claude 4.7 Opus shipped 2026-04-15 with a SWE-bench Verified delta of +4.9 points over Claude 4.6 Sonnet — Opus at 82.4, Sonnet at 77.5. That delta is the kind of jump that justifies a larger applied-research org behind the next training run, not a smaller one. A ten-year lease is what you sign when you expect the org chart that produces the next +4.9 point delta to live in the same building three years from now. The bridge read does not survive that assumption.
The tell for this read is the absence of leaked tour activity through 2027. Silence is the confirming evidence. If brokers in the South of Market submarket are not gossiping about Anthropic walk-throughs by the end of Q3 2027, the decade interpretation is the residual.
Scenario 3: The Optionality Reader (5-Year Anchor with Two 5-Year Extension Options)
The middle read is the one that gets the least press but is structurally the most common for late-stage AI labs at Anthropic's revenue trajectory. Picture a five-year anchor term with two five-year extension options at fair market rent, struck after the third year of the primary term and exercisable with twelve months' notice. The optionality is the point. The company commits to a footprint it can definitely fill for five years, retains the right to stay for fifteen, and pays a small premium on the headline rent for the option structure.
This is the read that survives the most scrutiny because it splits the difference between the bridge logic of Scenario 1 and the lock-in logic of Scenario 2. Five years is long enough to justify build-out spend — the TI allowance amortizes over a sixty-month term, not a twenty-four month one — and short enough that the company is not committing to a single building through whatever the AI industry's 2031 org chart looks like. The extension options are a hedge against the inverse: if the SOMA submarket loosens by 2029, the company gets to walk; if it tightens, the company gets to stay at a pre-negotiated framework.
For a tenant pricing Claude 4.6 Sonnet at $3 input and $15 output per million tokens, and the Opus tier at five times those numbers, the unit economics of the office footprint are subordinate to the unit economics of compute. A five-plus-five-plus-five lease lets the company keep real estate as a strategic asset rather than a strategic liability. The optionality is what large infrastructure-sensitive tenants buy when their own revenue model has more than one credible path through the next decade.
The implication for the eighteen months after April is more boring than either alternative. Hiring proceeds as planned. No second large lease is signed. Broker chatter stays quiet. The next funding round treats real estate as a fixed cost line that ticks up modestly per annum at the three percent escalator, with the embedded option valued at zero on the balance sheet because the company has not yet exercised anything. The tell for this read is also silence — but a different kind of silence than Scenario 2. It is the silence of a company that signed the right contract and has nothing to renegotiate for half a decade.
What All Three Share
Three things travel across all three reads, and they are the reason the lease term matters even though no one outside the company knows what it actually is.
First: the lease is being read as a proxy for compute and headcount projection, not for office strategy. No one cares about the carpet. The question is whether the contract implies Anthropic expects to be twice its current size in 2028 or four times. Every read above is really an argument about that growth multiple, with the lease term as the visible artifact. The shingle of journalism around the question is not real estate journalism. It is AI-trajectory journalism using a real estate document as the wedge.
Second: all three reads price the company's existing model release cadence into their assumptions. Claude 4.6 Sonnet on 2026-03-10. Claude 4.7 Opus on 2026-04-15. A frontier release every five to six weeks at this point in the cycle. That cadence requires a training org, an inference org, a safety org, a platform org, and a customer-facing org that are all hiring in parallel. The lease term has to seat that growth in some configuration. The three reads disagree on the shape, not the scale.
Third: none of the three reads are sourced to the document itself. Every number in circulation — the 24 months, the ten years, the five-plus-five-plus-five — is reconstructed from broker familiarity with the submarket, from analogues at peer companies, and from inference about what the sublessor would have agreed to. The article that finally gets the lease term right will be the one where someone files the underlying agreement with a regulator or a court. Until then, the question is open in the literal sense.
Which Scenario Is You
The reader of this piece falls into one of three groups, and the right interpretation depends on which group you are in.
If you are a competitor's strategy desk trying to model Anthropic's expansion velocity, Scenario 1 is the one to stress-test against. The 24-month read is the worst-case for your own real estate planning, because it implies a much larger second move is coming and you will be bidding against the same demand pool for the same submarket inventory in 2027.
If you are a recruiter trying to forecast the in-region hiring pipeline, Scenario 2 is the one to plan for. A ten-year lease is the strongest possible signal that the company expects to staff aggressively in the city, and the implied seat count gives you a ceiling for how much technical talent will be absorbed locally.
If you are an investor pricing the next round and trying to understand the company's capital allocation discipline, Scenario 3 is the one most consistent with the rest of the company's behavior. A five-plus-five-plus-five structure is what a tenant signs when they have institutional-grade legal advice and a CFO who values optionality over commitment. It is also the read that requires the fewest assumptions about the company's 2031 trajectory.
The lease term is a small number. The decision it informs is not.
A final note. The number to remember is not 24, ten, or fifteen. It is $75 per million output tokens on Claude 4.7 Opus. That number is the one that decides how much office the company can afford in any of the three scenarios. The lease will be sized to the inference business that price implies. The math is closed.
FAQ
What term lengths have actually been reported for Anthropic's SpaceX sublease?
Public reporting has put at least three durations in circulation: a short-term 24-month sublease, a decade-plus primary-term mirror, and a five-year anchor with renewal options. None of the three are sourced to a filed lease document. Each is reconstructed from broker familiarity with the SOMA submarket and from analogues at peer companies. The actual contract has not been published, which is why the divergence persists.
Why does the lease term matter for understanding Anthropic as a company?
The lease term functions as a proxy for the company's own headcount and compute projections. A shorter term implies the current footprint is a bridge to a larger deal; a longer term implies confidence that the current footprint plus its growth tranches will absorb expected hiring through 2031. The question is not really about real estate. It is about how aggressively Anthropic expects to scale the org behind Claude 4.7 Opus and its successors.
What does Claude 4.7 Opus tell us about the company's growth trajectory?
The April 2026 release shipped at 82.4 percent on SWE-bench Verified, 91.2 percent on GPQA Diamond, and 94.0 percent on HumanEval — a +4.9 point SWE-bench jump over Claude 4.6 Sonnet's 77.5. That cadence of frontier releases implies a training and applied-research org that is hiring continuously, not in steps. Whatever lease term Anthropic signed, the footprint has to seat that growth.
How does SOMA Class-A sublease pricing inform the question?
South of Market Class-A sublease pricing in late 2025 ran roughly $70 to $95 per square foot full-service for sub-three-year terms, against direct-deal pricing of $100 to $130 in the same buildings. A short sublease saves on rent and skips the build-out, which is consistent with the 24-month read. A longer sublease structured as a primary-term mirror would carry pricing closer to the direct-deal band. The pricing band is itself a tell about which read fits.
What signals would confirm or refute each scenario in the next eighteen months?
The short-term read is confirmed by a leaked second large lease signing in 2026 Q4 or 2027 Q1. The decade read is confirmed by absence of any such leak through end of Q3 2027. The five-plus-five-plus-five read is the residual: visible build-out spend without a second deal. Watch the broker gossip volume in SOMA. Silence and noise carry opposite meanings depending on the timing.
Why are AI labs particularly hard to model on real estate decisions?
The unit economics of compute dominate every other line item, including office. A tenant pricing inference at $75 per million output tokens is making decisions about model deployment first and seat count second. Real estate is downstream of the inference revenue forecast, not the other way around. This is why analogues from non-AI tenants in the same submarket break down quickly.
Is the lease term likely to ever be confirmed publicly?
The most common path to confirmation is a regulatory filing — most often a debt-covenant disclosure or a securities filing tied to a future funding round — or a court document if a dispute surfaces. Voluntary disclosure is uncommon for sublease structures of this size. The reader should expect the question to remain partially open until one of those mechanisms triggers, which may not happen inside the next eighteen months.
What would change if Anthropic's release cadence slowed?
A slower cadence — say, one frontier release per quarter instead of one every five to six weeks — would shrink the implied hiring trajectory and make the short-term lease read more plausible. The cadence is the load-bearing assumption underneath every scenario. If the next two releases stretch past their expected windows, revisit the bridge read first.