Perceptron Mk1.5 Brings Structured Spatial Output to Embodied Agents — and Retires Isaac the Same Day

Perceptron Mk1.5 is now generally available, and the interesting thing about it is not the benchmark table — it is what comes back in the response body. Ask a normal vision-language model where the forklift is and you get a sentence. Ask Perceptron Mk1.5 on a video frame and, alongside its prose, you can get machine-parseable geometry: a point, a bounding box, a polygon, a clip, or a <track> element carrying timestamped spatial observations across the whole file. That is the difference between a model that describes a scene and a model a robot controller can consume without a second parsing stage. It is the direct successor to Perceptron Mk1, the company's first release in May 2026, and it shipped on September 25 alongside the retirement of the Isaac 0.1 and 0.2 checkpoints that preceded it.

What Mk1.5 actually returns

The model is an embodied reasoning system for physical agents. On the input side it takes text, images, video and audio. On the output side it produces text plus optional structured annotations: points, boxes, polygons, clips and tracks. The tracking output is the piece worth dwelling on. Perceptron's docs describe a <track> block whose entries carry both a spatial observation and the timestamp it belongs to, so a 60-second clip comes back as a sequence of object positions over time rather than a single averaged guess about the scene.

A second detail that matters for anyone wiring this into a pipeline with more than one asset: Mk1.5 supports an asset_idx field, so a single request can reference multiple images or videos and address them separately. If your job is comparing a reference photo against a live camera frame — a defect-checking loop, an assembly verification step — that belongs in one call, not two.

The model also supports constrained responses, both JSON Schema and regex, which is how you turn "roughly there" into a schema your downstream code can validate. Function calling is supported on chat completions, and Perceptron's hosted MCP server now defaults to perceptron-mk1.5; passing model: "perceptron-mk1" explicitly is how you pin the previous default.

The spec sheet, and what it costs

The numbers here are worth stating plainly because they are modest in places a reader might not expect. The context window is 36,864 tokens — not a million, not 256K. Maximum output is 8,192 tokens. This is not a model you hand a two-hour video and a 300-page manual. It is sized for a tight perception task with a structured answer.

Pricing is $0.15 per million input tokens and $1.50 per million output tokens, with cached input at $0.0375 per million — exactly a quarter of the standard input rate. The client library is pip install "perceptron>=0.4.0", the endpoint is https://api.perceptron.inc/v1/chat/completions, and the key lives in PERCEPTRON_API_KEY. Nothing about the integration is exotic.

• Context — 36,864 tokens, versus the 32K window Perceptron Mk1 shipped with
• Max output — 8,192 tokens
• Input — text, images, video, audio (WAV, MP3, FLAC)
• Cached input — $0.0375/M, one quarter of the $0.15/M standard input rate
• Output — $1.50/M
• Reasoning control — reasoning_effort at high, medium, low, minimal or none, defaulting to high

That last row is a breaking change in disguise. Mk1.5 replaces the old vision_config.enable_thinking boolean with reasoning_effort, so any request you built against the previous model needs editing rather than merely repointing. The default of high is worth noticing for a different reason: if you do not set the field, you are buying the most expensive reasoning path on every call.

Audio, and video that carries its own soundtrack

Audio input is genuinely new here. Perceptron accepts it three ways — inline input_audio, an audio_url, or an audio_file_id — with a ceiling of 16,384 audio tokens per item. The docs put that at roughly 21.8 minutes of speech at about 750 tokens per minute, which is a reasonable budget for a shift-handover recording or a maintenance log.

More unusual is the video path. By default, Mk1.5 reads video as frames and ignores the audio track. Setting vision_config.enable_audio_in_video: true turns the soundtrack back on, which is the setting you want for anything where the noise is the signal — a machine that sounds wrong before it looks wrong, a spoken instruction issued off-camera, an alarm in the background of a site walkthrough. It is off by default, so a video with an audible fault will be silently analysed as a silent film unless you ask otherwise.

The benchmark picture, with the sourcing made explicit

Every figure below comes from Perceptron's own announcement and was measured by Perceptron. There is no Artificial Analysis index entry and no arena score for either Mk1.5 or its predecessor, so there is no third-party number anywhere in this comparison to set against them. Treat the figures as a vendor's claim about its own product, not as a neutral result.

On egocentric hand tracking the gap is wide and specific: Mk1.5 scores 0.9433 on hand_box against 0.4467 for Gemini 3.1 Pro and 0.6179 for the best Gemini in Perceptron's run, which the announcement identifies as Gemini 3.8 Flash. That is the +111% and +53% the launch post leads with, and it is the strongest single number in the release.

On general egocentric video understanding the picture is much closer. Perceptron reports EgoSchema at 80.40 for Mk1.5 against 81.20 for Gemini 3.8 Flash — a loss of 0.80 points — while claiming a 12% edge on the EgoSchema-hard subset at 63.75. A model that wins decisively on fine-grained hand geometry and loses narrowly on the broad comprehension benchmark is a specific kind of tool, and it is being sold as one.

Video object segmentation lands at 0.647 center-F1 and 0.628 point-HOTA on Molmo2-Track. Perceptron says that leads Molmo2-Track, Ref-DAVIS17 and ReasonVOS, while trailing MolmoPoint-8B on MeViS valid_u. Against the Qwen vision line the tracking comparison is worth reading carefully: the announcement lists Qwen3-VL-8B at 0.180 center-F1 from its paper, and Qwen3.5-27B at 0.530 and 0.476 — different checkpoints, different evaluation setups, and a paper number sitting in the same column as measured ones. That is not a like-for-like row.

Audio results are reported as DailyOmni 74.67, WorldSense 50.32, OmniBench 51.05 and AVHBench 80.56, with no comparison row attached. On multimodal search with tools enabled, Mk1.5 posts 56.0 on LiveVQA-W from a four-sample majority vote, against 53.2 for Gemini 3.8 Flash with Google Search grounding, 51.0 for GPT-6 sol with OpenAI web search, and 33.2 for GPT-5.2 online. Perceptron also claims a 36.1-point gain on MMSearch once tools are switched on. Majority voting over four samples is a legitimate technique and it flatters the score relative to a single greedy pass, so the 56.0 and the 53.2 are not measured the same way.

Latency, and how the 4.7× was measured

The speed claim is the one most likely to be quoted without its context, so here is the context. Perceptron reports up to 4.7× faster end-to-end from a median of three runs on a single H100, using LMArena prompts capped at 256-token answers, plus MIA-Bench and Video-MME with Perception Test. The 4.7× is the chat case: 5.2 seconds down to 1.1. Image question answering goes 1.7 seconds to 0.51, which is about 3.3×. A 60-second video processed eight at a time goes 19 seconds to 9.1, roughly 2.1×.

So the headline multiple is the best of the three, not the typical case. On a single H100, on short answers, the chat path really is 4.7× faster. On the video workload an embodied agent would actually run, it is closer to 2×. Both are good numbers; only one of them is the headline.

Isaac 0.1 and 0.2 were retired the same day

The Mk1.5 changelog carries a second entry dated September 25, and it is the one with teeth for anyone already in production. The isaac-0.1, isaac-0.2-1b and isaac-0.2-2b-preview model IDs were retired from the Perceptron API. They no longer appear in GET /v1/models, they are gone from the hosted MCP server, and requests naming them now fail with HTTP 404 model_not_found.

There is a second behavioural change buried in the same entry that will bite code which never mentioned the Isaac models at all: unknown model IDs now return 404 instead of the previous 400. Any retry logic, error branch or alerting rule that matched on a 400 for a bad model name is now matching on nothing. The migration path Perceptron gives is perceptron-mk1.5.

Retiring a model family on the same day you ship its replacement is a clean migration story and a harsh one. If you pinned Isaac because it worked, you have a deadline that already passed.

Where to run it, and how to try it without betting on it

Availability is through the vendor's own API and several third-party platforms. OrcaRouter does not currently route Perceptron Mk1.5 — the model is not in our catalogue — so the honest guidance is to go direct to api.perceptron.inc for it, which is exactly what the SDK and the PERCEPTRON_API_KEY environment variable are for.

What is worth saying anyway, because it applies to the decision rather than the model: a two-day-old checkpoint on a 36,864-token window with a reasoning default set to high is precisely the profile of something you should not put on a production path sight-unseen. Run it against your own frames first, and measure two things specifically — whether the structured outputs survive your actual edge cases, and what reasoning_effort: "high" costs you per call versus dropping to medium or low. The second is a one-line change with a direct effect on your bill, and it is the cheapest experiment in this release.

The broader pattern this fits — perception models returning geometry rather than adjectives — is one OrcaRouter is built to absorb. One API covers 200-plus models with provider list prices passed through at 0% markup, so a vendor price change on any of them is live on our side the same day, and automatic failover means a perception call can fall back to a second model instead of failing the request. If Mk1.5 is the only thing that clears your accuracy bar, none of that helps you today. If it is one candidate among several, running the comparison through a single endpoint is cheaper than wiring up a second SDK to find out.

What this release is and is not

Perceptron Mk1.5 is a focused tool with an unusually honest set of limits attached to it. It returns coordinates, not just description. It reads audio, including the soundtrack of a video if you switch it on. It is fast on short chat answers. It is also a 36,864-token model with an 8,192-token output ceiling, priced at $0.15 and $1.50, benchmarked entirely by its own vendor, and sold by a company that retired the previous generation in the same changelog entry.

If your problem is "find the hand, track it across the clip, tell me where it went and when," the hand-box number is the one to test. If your problem is broad video comprehension against a frontier generalist, the EgoSchema line — 80.40 against 81.20 — suggests you are buying a specialist at roughly parity, and the argument for switching has to come from the structured output and the latency, not from the accuracy.