Why did Harvey's gross margin go negative?

Harvey sells fixed annual seats with unlimited usage, but buys model inference by the token. After a March 2026 agent update, customer usage surged: Harvey reports a roughly twentyfold rise in token consumption in 2026 while annual recurring revenue only about doubled, from $190 million to more than $400 million. Bloomberg reported that gross margin fell from about 50% at the start of the year to roughly -50% by June 2026, meaning every revenue dollar cost about $1.50 in model bills alone. Agents that run unattended overnight through data rooms broke the ratio that per-seat pricing depends on.

What is Harvey Tenet?

Harvey Tenet is Harvey's in-house model, launched in August 2026. Harvey's engineering blog describes it as a Kimi K3 base, Moonshot AI's open-weight model, post-trained together with Fireworks research. Harvey says Tenet delivers better answers on Review Tables at roughly one-tenth the cost per cell and cuts Firm Knowledge cost per query by 90%. On Fireworks' Legal Agent Benchmark, Tenet costs $5.92 per task versus $5.62 for base Kimi K3 but lifts the all-pass rate from 10.8% to 19.7%, so the cost per fully passed task falls from about $52 to about $30.

Did Harvey raise prices to fix its margins?

No repricing has been reported. Harvey's margin turned positive again in August 2026 on the cost side, after moving flagship workloads onto Tenet. Harvey closed a $550 million round at a $15.6 billion valuation in September 2026 with more than 3,000 customer organizations. Artificial Lawyer reported in July 2026 that Harvey is understood to be developing a new pricing model, so metering may be deferred rather than ruled out, but the fix that actually landed was a cheaper model, not a more expensive seat.

Which other startups are moving to open-weight models?

Bloomberg named Abridge (clinical documentation), Decagon (customer support), Ramp (finance) and others as pursuing or weighing similar in-house open-weight pivots, with backing from investors including Sequoia Capital and General Catalyst. Rogo and Canva have also been reported as evaluating the move. The pattern is the same in each case: a per-seat or flat-priced agentic product whose token consumption grew far faster than revenue once agents ran autonomously.

Quick Answer

Why Harvey Swapped Claude for Kimi K3: -50% Margins

Published: September 24, 2026

The short answer

Harvey’s agents used twenty times more tokens in 2026 while revenue only doubled, its gross margin fell from about 50% to about -50% by June, and it fixed the problem by changing the model rather than the price. In August 2026 Harvey launched Harvey Tenet, a post-trained version of Moonshot’s open-weight Kimi K3 built with Fireworks, and margins turned positive again. Bloomberg’s September 21, 2026 report is the first hard number on what autonomous agents do to a flat-priced AI application, and it is why Abridge, Decagon and Ramp are reportedly weighing the same move. It is also, in the view of many observers, why Anthropic and OpenAI cut flagship prices within hours of each other the next day.

The arithmetic

	Start of 2026	June 2026	August 2026
Gross margin	~50%	~-50%	Positive again
Cost to serve per revenue dollar	~$0.50	~$1.50	Below $1.00
ARR	$190M	—	>$400M (September)
Token usage	Baseline	Rising sharply after March agent update	~20x year-to-date
Primary model	Frontier labs (Anthropic, OpenAI)	Frontier labs	Harvey Tenet (Kimi K3 post-trained)

The mechanism is simple. Harvey bills per seat with unlimited usage; Metronome’s pricing index describes it exactly that way. A partner who runs one query a week and an associate who leaves a diligence agent grinding through a data room overnight pay the same. Harvey’s model providers bill per token. Once the March agent update made unattended, long-horizon work the default, the metered input outran the fixed price by an order of magnitude.

For scale: Harvey started the year well ahead of its cohort. Bessemer’s 2025 benchmark put the fastest-growing AI “Supernovas” at roughly 25% gross margin, often negative. A 50% margin company going to -50% in six months is what an agent update looks like on a P&L.

What Harvey built instead

Harvey Tenet is “a Kimi K3 base that we post-trained together with Fireworks research,” per Harvey’s August 20, 2026 engineering post. The stated motive is blunt: open-weight models have cheaper per-token prices. Kimi K3 lists at $3/$15 per million tokens on Fireworks, against $5/$25 for Claude Opus 5 at the time and $10/$50 for Claude Fable 5.1 and GPT-6 Astra.

But a 40% token discount cannot produce Harvey’s claimed 90% cost cut on Firm Knowledge queries or one-tenth cost per Review Tables cell. The rest comes from three places:

Post-training for fewer steps. Fireworks’ own write-up shows Tenet costing slightly more per Legal Agent Benchmark task than base Kimi K3 ($5.92 vs $5.62) but nearly doubling the all-pass rate (10.8% to 19.7%). Cost per completed task drops from about $52 to about $30. That is the unit Harvey now optimizes.
Routing. A June 2026 Harvey–Fireworks study on a 100-task slice had open-weight GLM 5.1 do the work and call Claude Opus 4.7 as an “advisor” 0.83 times per task. The hybrid fully passed 18 tasks for $368; Opus alone passed 14 for $954. The frontier model becomes a callable tool, not the dependency the product sits on.
Capacity terms at Harvey’s volume that nobody outside the company can see.

The honest caveats

Every quality number in this story is Harvey’s or Fireworks’, measured on Harvey’s own benchmark, and the absolute pass rates are low: a 19.7% all-pass rate means four in five multi-step legal tasks still fail at least one check. Tenet is cheaper per completed task, not a frontier-beating model. And the provenance question we covered in August, of whose weights sit under privileged client documents, does not go away because the margin recovered; Harvey’s $550 million round in September was explicitly pitched around a model that keeps client documents under firm control.

Why the labs reacted

On September 22, 2026, one day after the Bloomberg story, Anthropic shipped Claude Opus 5.5 at $4/$20 with cache reads cut to $0.20, and OpenAI released GPT-6 Sol at $2/$10 and GPT-6 Luna at $0.10/$0.50. Both labs framed the cuts as efficiency gains. The timing suggests something else: a $15.6 billion customer publicly explaining that frontier pricing made its product unprofitable is the retention problem the price cuts are meant to solve. If Harvey can post-train Kimi K3 and leave, so can Abridge, Decagon and Ramp.

What it means if you buy or build agentic software

If you renew a per-seat agentic product (legal, clinical documentation, customer support, finance research), your vendor runs on the same arithmetic and has three exits: meter your usage, cap it, or change the model doing the work. Most contracts negotiate the first and ignore the other two. Ask which model serves your workload today, what changed in the last quarter, and for cost-per-completed-task reporting rather than token counts.

If you build one, the Harvey pattern is now the reference architecture: an open-weight base post-trained on your domain, a frontier model as an occasional advisor, and pricing that will eventually have to meter something. See Kimi K3 vs Claude Sonnet 5 on coding API cost for the raw price gap and How to cut AI API costs: caching, batching, routing for the routing half.

Last verified: September 24, 2026.