Treat Koa Decisions as a routing upgrade you will inherit inside Agentforce, and hold any plan to build on it until Salesforce publishes a price, an API date and accuracy for the model it ships. Salesforce announced Koa Decisions on November 6, 2026 as the next version of HyperClassifier, the model that picks which part of an agent handles a request. The launch post contains no accuracy figure. Its 26 millisecond latency number is the one Salesforce published in July for HyperClassifier. The research paper it points to reached its headline 85.5% on a GPT-OSS-20B checkpoint, while the post says the product is built on Google's Gemma.
If you run Agentforce, your job this month is to measure your current routing so you can tell whether the swap helped. If you need a decision model outside Agentforce, you cannot buy this one yet.
What Salesforce Announced and What You Can Use Today
Koa Decisions is a routing model you can demo today, with a developer API and builder tools still on the roadmap. A decision model reads an input and a fixed list of options, then returns one choice, a yes or no, or a score in a single step instead of generating text. Salesforce says it built Koa Decisions by applying "novel training techniques" to Gemma, and that it returns a confidence score with each answer. It shares a family name with the Koa reasoning model built with NVIDIA, which is a separate model with a separate job.
The post says "this month, we're introducing" the model. It does not say whether Koa Decisions is already answering in customer orgs, whether admins have to do anything, or when the change reaches a given region. Its footnote says availability "may vary by region and is governed by customer agreements." Flows, Apex, prompt templates and a developer API are described as plans, with no dates, and the post gives no price.
The demo is the one thing you can touch. Salesforce says a free Trailhead account is enough, including for non-customers, and the page itself asks you to "choose a Salesforce org with access to Koa Decisions."
You are probably on its predecessor already. HyperClassifier became generally available in Spring '26 and is the default intent detection and routing model for Agentforce Service, and the same model handles semantic endpointing in Agentforce Voice, the check for whether a caller has finished speaking.
Which Latency Number Should You Plan Around?
Plan around the half second until Salesforce states how the 26 millisecond figure was measured on the shipping model. Salesforce has now published three different speeds for this layer of Agentforce.
| Source | Date | Model named | Figure | Stated conditions |
|---|---|---|---|---|
| Right-sizing post, body | July 8, 2026 | HyperClassifier | About 26 ms vs. about 1,446 ms for a general frontier model | None |
| Right-sizing post, table | July 8, 2026 | HyperClassifier | About 50 ms vs. roughly a second | None |
| arXiv preprint v2 | October 10, 2026 | Koa-decision (checkpoint unstated) | 0.53 s median, 0.63 s p95 | End to end, one request at a time, no batching, no client-side caching |
| Koa Decisions launch post | November 6, 2026 | "our decision model" | Roughly 26 ms vs. about 1.4 s for a general-purpose LLM | Routing requests in Agentforce; no percentile, load or hardware |
The launch post never says its 26 milliseconds was measured on Koa Decisions, and the matching July figure was explicitly HyperClassifier's. Unite.AI reached the same reading, noting that the measurement "belongs to the predecessor."
The numbers can all be honest. The paper times the whole trip through a shared serving interface. Salesforce's engineering account of HyperClassifier describes caching the repeated instructions and cutting network hops, and the paper's own batch study puts per-sample median latency for a Gemma-3-27B checkpoint between 5.6 and 54.5 milliseconds. A gap between model time and serving time is the likely explanation, but that is an inference, since Salesforce gives no percentile, load or hardware for the short figures.
The half second is still good. In the paper, it is about 1.9 times faster than the quickest frontier model tested (GPT-4.1-mini) and about 7.5 times faster than the slowest (Gemini-Pro-3.1), and the p95 sits only a tenth of a second above the median. Against Jev, the comparable single-token system, the margin is 0.53 seconds to 0.55 at the median on the voice test, too small for a caller to notice. Single-token decoding explains most of the speed, and Jev has it too.
Does the 85.5% Apply to the Model You Will Get?
Not as published, because the paper's best intent-routing score belongs to a different base model from the one Salesforce says it shipped. The paper's abstract reports 85.5%. Table 2 shows that row is GPT-OSS-20B, the base model Salesforce named for HyperClassifier in July. The Gemma checkpoint in the same table scored 81.43%.
| Benchmark (accuracy %) | GPT-OSS-20B | Gemma-4-26B | Jev | Best frontier model tested |
|---|---|---|---|---|
| Production intent routing (587 utterances) | 85.52 | 81.43 | Not run | 87.73 (Claude-4.8-Opus) |
| Voice endpointing (353, binary) | 96.90 | 94.33 | 93.20 | 97.17 (Claude-4-Sonnet) |
| SST-2 sentiment (872) | 92.89 | 95.76 | 96.10 | 96.79 (Gemini-Pro-3.1) |
| Amazon Reviews (5,000) | 95.34 | 96.60 | 96.80 | 97.42 (Gemini-Pro-3.1) |
At 81.43%, the Gemma-4-26B checkpoint trails every frontier model on the panel except GPT-4.1-mini. A different Gemma checkpoint, Gemma-3-27B, reached 84.7% on a held-out intent-routing set in the paper's appendix, which gives no utterance count, so Gemma may not be the problem in itself. The paper calls the method Koa-decision and does not say which checkpoint is the product. Until Salesforce says so, any of these numbers could be the one that applies to you, or none of them.
On a 10,000-conversation month, 85.5% means about 1,450 misrouted requests and 81.4% means about 1,860. The routing benchmark also has limits a buyer should weigh. It is 587 utterances from one production agent, and Salesforce says it cannot release the data, so nobody outside the company can reproduce the result. A simple binomial estimate on 587 samples puts the margin at roughly three points either way, which covers the distance between GPT-OSS-20B and the best frontier scores. Jev was not run on this set, so the paper's claim that Koa-decision is "competitive" with Jev rests on public sentiment sets, where Jev beat both checkpoints, and a voice test, where both checkpoints beat Jev.
Give Salesforce credit for the disclosure. A table with every backbone, p95 latency and a losing row is more than Microsoft printed for Decision-1. The launch post simply left all of it out.
Image, Audio and Confidence Scores Have the Thinnest Evidence
Audio has no published evaluation at all. The launch post says Koa Decisions "adds multimodal support for images and audio," while the paper defines its inputs as text and vision and lists going beyond them as future work. Its voice endpointing test runs on text transcripts.
Vision has one benchmark. On MIntRec 2.0, a 30-class test of intent in video dialogue, a fine-tuned Gemma-3-27B scored 62.72%, ahead of Gemini-Pro-3.1 at 55.93% and two task-specific encoder models at about 60.6%. That is the best score on the board, and it still gets more than one in three wrong. Expect a damaged-product photo routed to the wrong queue at a rate you should measure before you trust it.
Three more limits are stated in the paper. The training corpus is entirely English, and on the six-language MTOP test, Gemma-4-26B averaged 73.67% to GPT-4o's 77.70%. Multi-label output appears in the abstract, but the paper omits how it is built and reports no results for it. And the authors write that "the intended use is intent routing, not user profiling or high-stakes automated decisions," a limit worth rereading when the same model reaches flows and Apex, where a decision can approve a refund or close a case.
The confidence score needs its own test. The demo walkthrough in the launch post shows routes at 95% and 99.5%. The paper reports only accuracy and latency, with no calibration measurement. When Artificial Analysis charted calibration for nine decision models on October 13, accuracy fell below stated probability "for all decision models" at almost every level. Koa Decisions was not in that test. Use the score to rank answers, and do not set an automation threshold on it until you have checked it against your own labels.
How Koa Decisions Compares With Models You Can Buy
Three alternatives have a published price and an API today, so Koa Decisions only wins if your decisions already live inside Agentforce. Prices below were checked on each vendor's page on November 8, 2026.
| Koa Decisions | TypeSafe Jev 1.14 | OpenAI Decisions API | Microsoft Decision-1 | |
|---|---|---|---|---|
| Status | Announced; Agentforce routing, demo | Available | Generally available | Foundry and OpenRouter |
| Input price per million tokens | Not published | $0.042 | $0.10 uncached | $0.042 |
| Output tokens | Not published | Free | Free | Free |
| Image input | Claimed | Yes | Yes | Not stated in launch post |
| Audio input | Claimed, no evaluation | No | Not supported as input | Not stated in launch post |
| Developer API | Planned, no date | Yes | Yes | Yes |
| Base model | Gemma (version unstated) | Not disclosed | GPT-6 Luna | Qwen3.5-9B |
Do not let the token price drive this decision. At an assumed 2,000 input tokens per decision, a million decisions costs $84 on Jev and $200 on OpenAI. A standard Agentforce action is $0.10, so a single action costs about 1,200 times what one Jev decision does at that size. The November 3 Flex Credits rate card has no row for routing, classification or decision models; its only per-decision line sits under Personalization. Your Agentforce bill follows the actions an agent takes, and a better router pays off through fewer wrong actions and shorter waits.
The price question opens up when Koa Decisions reaches flows and Apex. The same rate card no longer carries the "Salesforce Process Invocation (Flows, Apex)" line that the August 18 version listed as "TBA," and its one Headless 360 usage type, "Headless Platform Interaction," is also "TBA," so a builder who designs a decision step into a flow today has no number to put in the business case.
Keep your existing classifiers for now. Einstein Case Classification trains on your own closed cases, needing at least 400 and ideally 1,000 or more, while a zero-shot router works from label descriptions alone. Nothing Salesforce has published compares the two on the same data, and a fair test on your cases is cheap. Our decision model buyer's guide has the approach, including where a small fine-tune beat Jev.
What to Do Before You Rely on It
Build a labeled baseline now, because you cannot judge a swap you did not measure beforehand.
This Week:
- Ask your Salesforce account team in writing which model serves routing in each of your orgs today, whether and when Koa Decisions replaces it, in which regions, and whether you can pin the old model or roll back.
- Have your Agentforce admin export 500 recent utterances per high-volume agent, with the subagent each should have reached, labeled by someone who owns that queue.
- Run your 20 hardest cases through the demo, including any that involve images, and record the confidence score beside whether the answer was right.
This Month:
- Record routing accuracy per subagent and p50 and p95 routing time from your own session traces, so the swap has a before and after.
- Run the same labeled set through Jev and the OpenAI Decisions API with your existing subagent descriptions. Five hundred utterances at 2,000 tokens is one million tokens, which costs $0.042 on Jev and $0.10 on OpenAI.
- Rewrite any subagent description a new hire could not act on. The paper's zero-shot design decides from those descriptions, so vague ones become misroutes.
- If you route in languages other than English, split the test by language and treat each one as its own pass or fail.
Before You Build a Flow or Apex Step on It:
- Get the price and meter in writing, plus notice terms for any change of base model.
- Keep customer-facing refunds, eligibility calls and anything with legal exposure on a path with a human or a deterministic rule, in line with the paper's stated intended use.
- Retire an Einstein classifier only after it loses to the router on the same cases.
The Bottom Line
Salesforce has been swapping this layer for a while. It claimed a 30x speedup for topic classification in January, gave millisecond figures in July and announced the successor in November, and none of those posts gave a percentile, load or hardware behind the number. The method is sound, and independent systems like Jev show single-token decisions work. As with Clef against Jev, the published lead is narrow and task specific, and here it is attached to a checkpoint that may not be the one in your org.
Ask your account team which checkpoint answers your routing calls, and get the reply in writing.
Continue Reading
- Salesforce Koa vs Claude vs Gemini: Its Paper Only Beats GPT-4.1
- An 18-Minute Fine-Tune on a Gaming GPU Beat Jev on a Phishing Test
- Long Rubrics Can Make OpenAI Decisions API Cost 4x More
- Microsoft Brings Decision-1 to Foundry and OpenRouter at Jev's Price
- Your Agentforce Bill Depends on How Many Actions an Agent Takes
- Cloudflare's Clef Beats Jev at Routing and Loses at Judgment