![Samaya Sterling Scores 58.9% vs Claude’s 55.9% [2026]](https://shattered.io/wp-content/uploads/2026/10/samaya-sterling-3b-active-finance-agent-frontierfinance-58-9-percent-2026-gen.webp)
Samaya AI says its new Samaya Sterling model scored 58.9% on the FrontierFinance benchmark, ahead of Claude Fable 5.1 at 55.9%, while costing $0.61 per question instead of $6.89. The company posted the results on October 28, 2026, and MarkTechPost and nokiapoweruser.com have since carried the story. Sterling is a post-trained model for financial research, and Samaya describes it as built on NVIDIA Nemotron 3.5 Lightning, a 30B mixture-of-experts model with 3B active parameters.
The headline is a cost story more than a quality story. A three-point edge on a 220-question public benchmark is real but modest. A price gap of roughly 11x is the part that will push finance technology buyers to read the footnotes. This analysis walks through what Samaya published, what the arithmetic says, and which claims still need outside confirmation.
The release also fits a pattern we have tracked all month. Smaller, cheaper systems keep aiming at one narrow job and challenging frontier generalists on price. We saw it when Mistral’s Le Chonk coding agent matched a rival on a coding index, and Sterling applies the same playbook to finance. Our wider AI and machine learning coverage follows the same thread.
Don't miss new tech stories on Google
Add Shattered once in the Google app and our stories show up in your news suggestions.
What Samaya AI announced on October 28 #
Samaya published a research post dated October 28, 2026, alongside a release page for Sterling. The company opened its announcement with a plain statement of intent. Samaya AI wrote: “Introducing Samaya Sterling: our first post-trained model for state-of-the-art financial research.”
The word “first” matters. Sterling is the first model Samaya has post-trained itself for this purpose, which tells us the company is moving from wrapping other firms’ models toward owning a model of its own. That is a strategic shift, and it explains why the launch leans so hard on a cost comparison. A company that trains its own model can set its own price curve.
The second quoted line carries the commercial pitch. Samaya AI wrote that Sterling “outperforms Claude Fable 5.1 (58.9% vs. 55.9%) at 11× lower cost.” Every figure in that sentence traces back to Samaya’s own FrontierFinance evaluation. No independent lab has published a replication in the record we reviewed, so treat the claim as a vendor result until someone else reruns it.
One caution on wording. The word released in early headlines rests on Samaya’s own release and research pages. We found nothing from a third party confirming availability terms, pricing tiers for customers, or a general-access date. Where this article says Sterling launched or debuted, it refers to those Samaya publications and nothing more.
The FrontierFinance numbers, side by side #
FrontierFinance is the benchmark Samaya chose, and the company describes it as 220 public questions. Sterling earned 58.9% of rubric items. Claude Fable 5.1 earned 55.9%. The table below lays out every figure Samaya published, plus the gaps we calculated from them.
| Metric | Samaya Sterling | Claude Fable 5.1 | Gap (our math) |
|---|---|---|---|
| FrontierFinance score (rubric items) | 58.9% | 55.9% | 3.0 points |
| Cost per question | $0.61 | $6.89 | $6.28 cheaper, about 11.3x |
| Mean latency | 282 seconds | 268 seconds | 14 seconds slower, about 5.2% |
| Rubric items missed | 41.1% | 44.1% | 3.0 points fewer misses |
| Benchmark size | 220 public questions | 220 public questions | None |
Read the table from the bottom up. The test is public and small, which makes it easy to inspect and also easy to tune toward. The score gap is three points. The cost gap is more than tenfold. The latency gap runs against Sterling, though only slightly. Together those rows describe a model that trades a few seconds for a large price cut and a small accuracy bump.
Checking the 11x cost claim against the listed prices #
Samaya’s “11× lower cost” line is a rounded figure. The listed prices are $0.61 per question for Sterling and $6.89 for Claude Fable 5.1. Divide the second by the first and you get about 11.3. Samaya rounded down, which is the conservative direction, so the headline does not overstate its own data.
What the figure does not tell us is how Samaya measured cost. The announcement coverage we reviewed gives a per-question number but not the token counts, the tool calls, or the retry policy behind it. Research agents often make many calls per question, so a per-question cost folds in the agent loop, not just the model price. That is a fair way to compare, as long as both systems ran under the same loop. We cannot confirm that from the record.
To show why the gap matters, here is simple multiplication at several volumes. These are our projections from the two published per-question prices, not figures Samaya reported. They assume the mean cost holds as volume grows.
| Questions run | Sterling at $0.61 | Claude Fable 5.1 at $6.89 | Savings |
|---|---|---|---|
| 220 (one full FrontierFinance pass) | $134.20 | $1,515.80 | $1,381.60 |
| 1,000 | $610 | $6,890 | $6,280 |
| 10,000 | $6,100 | $68,900 | $62,800 |
| 100,000 | $61,000 | $689,000 | $628,000 |
For a research desk that fields a few thousand questions a month, the savings are a rounding error against analyst salaries. For a platform that runs hundreds of thousands of automated research tasks, the gap decides whether a product has a margin at all. That second group is the audience Samaya is courting.
Why 3B active parameters is the number to watch #
Sterling’s base model has 30B total parameters but only 3B active for any given token. That is 10% of the model doing the work on each step. In a mixture-of-experts design, a routing layer picks a small group of expert sub-networks per token, so the compute bill tracks the active count rather than the total.
This is the structural reason the price gap can exist at all. A dense frontier model spends its full parameter budget on every token. A sparse model with 3B active spends a fraction of that. Samaya did not publish a parameter count for Claude Fable 5.1, and we will not guess one. What we can say is that the architecture of the base model points toward cheap inference.
Cheap inference also changes who can host it. A model with 3B active parameters can run on far smaller hardware than a frontier system. The record does not say whether Samaya offers self-hosting, so we flag that as an open question rather than a feature. Teams that care about keeping research data in-house should ask the question directly.
Nemotron 3.5 Lightning as the foundation #
Samaya built Sterling on NVIDIA Nemotron 3.5 Lightning, which it describes as a 30B mixture-of-experts model with 3B active parameters. Starting from an open-style base and adding domain training is now a common route for specialist vendors. It skips the cost of pretraining from scratch and lets a small team spend its budget on the data and reward design that make a model good at one thing.
The Nemotron name has been busy on our desk. We recently reported that Nemotron 3.5 Transcribe fits in 17 GB with a 5.33% word error rate, a sign that the 3.5 family is built for efficient deployment. Sterling extends that family into a reasoning-heavy, text-based research task, which is a very different workload from speech recognition.
Samaya reports that post-training lifted performance on FrontierFinance from 42.5% to 58.9%. That is a gain of 16.4 points, or roughly 39% in relative terms. The base model alone would have trailed Claude Fable 5.1 by 13.4 points. So the story is mostly a post-training story. Nvidia supplied the engine, and Samaya’s training recipe moved it from also-ran to leader on this test.
The four-stage training recipe #
Samaya says it used a four-stage process, with two supervised-learning stages followed by two reinforcement-learning stages. The sequence is a familiar one. Supervised stages teach the model what good output looks like, and reinforcement stages push it toward higher rubric scores by rewarding better answers.
What the record confirms
The confirmed facts are narrow. There are four stages, split evenly between supervised and reinforcement learning. The end result raised the FrontierFinance score from 42.5% to 58.9%. Samaya presents the full process as the reason Sterling outperforms its base model.
What the record leaves open
The announcement coverage we reviewed does not describe the training data, the reward model, or the compute budget. It does not say how much of the lift came from each stage. Those details decide whether other teams can copy the approach, and whether the rubric used in reinforcement learning overlaps with the rubric used in evaluation. That last point is the one outside reviewers will probe first.
The latency trade-off: 282 seconds versus 268 #
Sterling’s mean latency is 282 seconds, against 268 seconds for Claude Fable 5.1. That is about 4 minutes 42 seconds versus 4 minutes 28 seconds. The 14-second difference works out to roughly 5.2% slower per question.
Both numbers sit in the range of minutes, not seconds. That tells us these are deep research tasks where the agent searches, reads, and writes over many steps. In that setting, a 14-second gap barely registers. An analyst waiting for a memo will not feel the difference between 268 and 282 seconds, but a pipeline running thousands of jobs in parallel will feel the cost difference immediately.
The latency figure also hints at a limit of the sparse design. Sterling is cheaper per token, yet it still takes slightly longer overall, which suggests it may use more steps or longer outputs to reach its answers. Samaya did not break that down. If you plan to deploy it for interactive use, test the tail latency, not just the mean.
What FrontierFinance does and does not prove #
FrontierFinance, as Samaya describes it, has 220 public questions scored against rubric items. Public questions bring transparency. Anyone can read them, rerun them, and check the scoring. They also bring risk, because public test sets can leak into training data and invite tuning toward the exact format of the rubric.
A 3.0-point lead on 220 questions is small enough that the margin could shrink under a different grader or a different prompt setup. We have no confidence intervals from Samaya, and we will not invent them. The honest reading is that Sterling is competitive with Claude Fable 5.1 on this test, and that the cost advantage is the firmer finding.
The verdict also has a scope limit. The claim that Sterling beats Claude Fable 5.1 holds for the cited FrontierFinance evaluation. It does not extend to coding, general reasoning, multilingual work, or other finance tasks outside the benchmark. Vertical benchmarks reward vertical training. A model tuned for one exam should do well on that exam.
Compare that with how generalist models get judged. When we looked at Gemini 4 Carbon’s index and SEC-bench results, the question was breadth across many tasks. Sterling’s pitch is the opposite, depth on a single professional workflow.
Claims scorecard: confirmed, rounded, and unproven #
The table below separates what the record supports from what it does not. We built it to help readers quote the story accurately.
| Claim | Status | Basis |
|---|---|---|
| Sterling scored 58.9% on FrontierFinance | Reported by Samaya | Samaya research post, October 28, 2026 |
| Claude Fable 5.1 scored 55.9% | Reported by Samaya | Same evaluation, run by Samaya |
| Sterling costs 11x less | Rounded | $6.89 divided by $0.61 is about 11.3 |
| Sterling released | Self-reported | Samaya release and research pages only |
| Sterling beats Claude broadly | Not established | Only the cited benchmark supports it |
| Independent replication | None found | No third-party rerun in the record |
| Weights, licensing, customer pricing | Not covered | Absent from the coverage we reviewed |
Who stands to gain: finance research teams and platforms #
Financial research is a natural target for agent models. The work is document-heavy, it follows repeatable structures, and it has clear quality rubrics. Analysts spend hours pulling facts from filings, transcripts, and market data, then writing them up. An agent that handles the first pass at $0.61 a question changes the economics of coverage.
The buyers who benefit most fall into two groups. Research platforms that serve many clients can run far more questions per dollar of compute. Smaller funds and advisory shops, which could never justify $6.89 a question at scale, get access to automated research they would otherwise skip. In both cases, the savings are real only if the quality holds up on their own questions, which may look nothing like FrontierFinance.
Risk teams will ask a different question. A cheaper model that is wrong with confidence can do more damage than an expensive one, because it gets used more. The same logic applies to security. Our piece on zero trust for AI agents and its three tiers after login argues that agents with access to sensitive data need tight permission controls, and a finance research agent is a clear example.
Specialist versus frontier generalist: the competitive picture #
Sterling pits a narrow, sparse, post-trained model against a general frontier system. Each side has a real advantage. The generalist handles any task and improves whenever its maker ships an update. The specialist wins on price and can be tuned relentlessly toward a single rubric.
Samaya’s result suggests the specialist route is viable even against a current top-tier model, at least on one test. If other vertical vendors reproduce the pattern in law, medicine, or accounting, frontier labs will face pressure on their per-token pricing for professional workloads. They have two obvious replies. They can cut prices, or they can ship their own domain-tuned variants.
Teams choosing between the two should run their own comparison. Our walkthrough of a ChatGPT, Claude, and Gemini API test harness in 12 steps shows how to score models on your own prompts, and the same method works for a specialist like Sterling. Vendor benchmarks tell you where to look. Your own data tells you what to buy.
Historical context: the shift from wrappers to owned models #
For the past few years, many AI startups built products by wrapping a third-party model with prompts and retrieval. That approach shipped fast, but it left margins at the mercy of API pricing. Every price cut by a frontier lab helped, and every price hike hurt. The next phase has been vendors taking an open or semi-open base model and post-training it for one domain. Sterling is a clean example, with Samaya calling it its first post-trained model. The move gives a company control over cost, latency, and behavior, at the price of owning a training pipeline.
The open-weight ecosystem made this possible. Base models now arrive in sizes and licenses that vertical vendors can build on. Our comparison of open-weight model licenses across six families shows how much the terms differ, and license terms will matter to anyone who wants to build on a base like Nemotron. Check them before you commit.
Market impact: a price signal for AI-heavy workloads #
The market reaction to model launches now runs through one question: who pays for the compute? Big technology companies have been spending heavily on AI infrastructure, and investors have grown sensitive to whether that spending earns a return. We tracked that mood in our report on Big Tech’s $750 billion AI bet meeting slowing core growth.
A credible 11x cost reduction for a high-value workload points the other way. It suggests that demand for professional AI work can be served with far less hardware per task than the frontier approach implies. That is good news for buyers and a quiet warning for anyone whose revenue depends on expensive tokens staying expensive.
One result will not move a sector. Samaya’s numbers have not been replicated, and the benchmark is small. Still, the direction matches a broader trend. Efficient, specialized models keep gaining ground on tasks where a rubric defines success, and finance has plenty of rubrics.
Five predictions for finance research models #
These are our own forecasts, based on the record above. None comes from Samaya or any named analyst.
- Independent reruns will land within weeks. A public 220-question benchmark invites replication. Expect outside groups to test Sterling and report scores that may differ from Samaya’s by a few points in either direction.
- Rival vertical vendors will publish similar cost-versus-score charts. The format is persuasive, and competitors will want their own version aimed at Claude Fable 5.1 or similar systems.
- Frontier labs will answer with pricing or domain variants. A gap of this size on a valuable workload gives them a reason to act, either by trimming prices or by releasing tuned finance offerings.
- Benchmark integrity will become the main fight. Questions about overlap between training rubrics and evaluation rubrics will grow louder as more specialist models post leaderboard wins.
- Buyers will move to hybrid setups. Many teams will route routine research to a cheap specialist and reserve a frontier model for hard or high-stakes questions.
How to evaluate Sterling before you trust it #
Any team considering Sterling should start with its own questions. Pull 50 to 100 real research tasks from your last quarter, write rubrics that match how your analysts grade work, and run both Sterling and your current model. Track score, cost, and latency together, because the three trade against each other.
Ask Samaya for specifics the announcement left out. Request the cost methodology, the data-handling terms, the availability and licensing details, and any third-party evaluation. Check how the model behaves on questions it cannot answer, since a confident wrong answer in finance is costly.
Keep a human reviewer in the loop for anything that reaches a client or a trade decision. A 58.9% rubric score means the model missed over 41% of the rubric items on Samaya’s own test. Claude Fable 5.1 missed even more, but neither system is ready to run unsupervised.
Frequently asked questions #
What is Samaya Sterling?
Samaya Sterling is a post-trained model for financial research from Samaya AI. Samaya calls it its first post-trained model and says it is built on NVIDIA Nemotron 3.5 Lightning.
How does Sterling compare with Claude Fable 5.1?
On Samaya’s FrontierFinance evaluation, Sterling scored 58.9% and Claude Fable 5.1 scored 55.9%. Sterling cost $0.61 per question against $6.89, and its mean latency was 282 seconds against 268.
Is the 11x cost claim accurate?
It is a rounded figure. Dividing $6.89 by $0.61 gives about 11.3, so the claim matches Samaya’s own listed prices. The cost methodology behind those prices was not detailed in the coverage we reviewed.
What is FrontierFinance?
Samaya describes FrontierFinance as a benchmark of 220 public questions, scored against rubric items. Sterling’s 58.9% refers to the share of those rubric items it satisfied.
What does 3B active parameters mean?
The base model has 30B total parameters in a mixture-of-experts design, but only 3B are active for each token. That keeps compute per token low compared with a dense model of similar total size.
How was Sterling trained?
Samaya says it used a four-stage process, with two supervised-learning stages and two reinforcement-learning stages. It reports that this raised the base model’s score from 42.5% to 58.9%.
Does Sterling beat Claude Fable 5.1 at everything?
No. The available record supports the claim only for the cited FrontierFinance evaluation. Broader superiority across other tasks has not been shown.
Has anyone independently verified the results?
Not in the record we reviewed. The figures come from Samaya’s own research post dated October 28, 2026, and the coverage by MarkTechPost and nokiapoweruser.com repeats those figures.