AussieBytes Blog ◖ᵔᴥᵔ◗

No wall, just a firehose: A crazy week of frontier models and what it means for enterprise

TL;DR

The last week in AI has been ridiculous if you care about models:

The pace isn’t slowing, as we discussed previously, the innovation pace has been unprecedented this year. Pre-training is still buying capability gains; post-training (reasoning modes, RL, tools, safety) is where vendors now differentiate.

For enterprises, that means:

1. The week that was: four labs, one vision model and a lot of churn

1.1 Google: Gemini 3 is top of the class (but still hallucinates)

Gemini 3 Pro is now Google’s flagship underpinning multiple products including Search AI Mode, the Gemini app/API, Vertex AI and the new Antigravity IDE. It sits at or near the top of public leaderboards and posts state-of-the-art scores on GPQA Diamond, MMMU-Pro and Video-MMMU, with genuinely useful video understanding in day-to-day workflows.

Gemini Pro 3

The catch is that on Artificial Analysis’ Omniscience Index, Gemini 3 Pro has top accuracy but around 88% of its non-correct answers are hallucinations – if it doesn’t know, it often still guesses. By contrast, OpenAI reports ~2.8% hallucinations on FActScore and ~1–2% on LongFact for GPT-5, and Vectara’s leaderboard has Claude Sonnet 4.5 at ~5.5% hallucinations on short-doc summarisation. These models are much better at saying they don't know.

Potential applications:

In practice, most teams may eventually end up with Gemini for multimodal/search and GPT-5/Claude for low-hallucination factual workloads, rather than betting on a single “winner", but the landscape is moving rapidly.

Takeaway: I am finding Gemini great for rich retrieval, UX; and sticking to OpenAI/Anthropic where being accurate/repeatable matters.”

1.2 OpenAI: GPT-5.1 & Codex-Max – long runs, reliable, with major potential

OpenAI’s week was about reasoning and agents:

codex OpenAI, Nov'25

METR offers a more grounded view: for long-horizon software work, Codex-Max has a 50% “time horizon” of ~2h42, up from ~2h17 for GPT-5 – part of a smooth curve where that horizon has roughly doubled every 4–7 months since the o-series. Day-long reliable agents are still a few doublings away.

metr codex max METR, Nov'25

For CTO's:

Google now leads on some most reasoning/multimodal benchmarks this week. Interestingly, reports claim that Gemini’s latest pre-training run “leapfrogged” GPT-5. But GPT-5/5.1 and Codex-Max still look best-in-class on factuality and long-horizon coding, which is what matters for most production use in the software development space.

Takeaway: In practice, initially I am leaning on GPT-5.1 (Thinking/Pro) for planning, GPT-5.1-Codex-Max for refactors and deep debugging, Gemini 3 for front-end and creative UX, and Claude Sonnet/Opus for security review and cyber risk work. But more testing over next few weeks / months will validate (or not) this approach.

1.3 xAI: Grok 4.1 – cheap, good EQ, use selectively

Grok 4.1 is xAI’s first truly frontier-class model. It ships in thinking and non-reasoning variants across grok.com, X and partner APIs, sits just behind Gemini 3 Pro on some LMArena views, and is tuned hard for emotional intelligence and creative writing.

Hallucinations are down from ~12% to ~4.2% for non-reasoning Grok 4.1, with FActScore errors around 3% – competitive with Claude and Gemini, still above GPT-5’s best factual scores. Pricing is the hook: public rates around US$0.20 / 0.50 per million input/output tokens in some channels are far below other frontier models.

grok 4

Sweet spots:

Takeaway: Grok 4.1 is great when you want personality and low cost - good also as a secondary model for specific use cases.

1.4 Anthropic: Expanding access to Claude models

Anthropic didn’t ship a fresh model this week, but Claude Sonnet 4.5 and Haiku 4.5 are now everywhere:

Sonnet 4.5 is great for coding and computer use (SWE-bench Verified, OSWorld, Terminal-Bench), in the same frontier tier as GPT-5.1 and Gemini 3 on general reasoning, with Gemini 3 Pro clearly ahead on the hardest multimodal and visual benchmarks.

Claude is among the most cloud-diversified frontier family, with a strong safety reputation and deep productivity-suite integrations, though its capacity and rate-limit policies are still evolving. If you’re already an M365 or AWS shop, Claude is a natural co-primary model alongside OpenAI or Gemini, especially for agents, spreadsheets and doc/slide drafting.

On safety, Anthropic is more forward-leaning in its model governance and catastrophic-risk frameworks than most labs, while OpenAI tends to be more conservative on day-to-day enterprise controls (data use, Agent Mode, tools). For CTOs, that means Anthropic is safety-forward in how it thinks about future models, but the practical execution risk from using Claude today is not clearly lower than with ChatGPT Enterprise.

Takeaway: Claude gives you reach and safety-forward model governance across clouds, but execution risk is not necessarily lower than OpenAI’s – you still need to manage agent security, rate limits and capacity as you scale.

1.5 Meta: SAM 3 – new vision capabilities

Meta’s contribution this week is a vision foundation model, a welcome distinction from the LLM focused week it was. Segment Anything Model 3 (SAM 3) is an 848M-parameter system that can detect, segment and track all instances of a concept in images and video based on short text prompts (“yellow school bus”, “player in a red jersey”) or visual examples.

sam 3 Sam 3 - Promptable Visual Segmentation

On Meta’s SA-Co benchmark (270k concepts, millions of instances), SAM 3 reaches about 75–80% of human performance and more than doubles prior open-vocabulary systems like OWLv2, DINO-X and even Gemini 2.5 on key detection metrics. A single model handles long video sequences and crowded scenes, maintaining consistent tracks across frames.

Meta also released SAM 3D, which reconstructs 3D objects and human bodies from a single image, useful for e-commerce spins, AR “view in room” experiences, digital twins and robotics. Both SAM 3 and SAM 3D ship with open weights and code, with an ecosystem forming around fine-tuning and annotation.

Enterprise use-cases:

Takeaway: The enterprise opportunity is not just language – vision and 3D offer major opportunities to rethink workflows.

2. Pre-training vs post-training: no wall yet

Taken together, this week’s launches show that both pre-training and post-training still have meaningful headroom:

Pre-training: what this week shows

Post-training & tools: where vendors really diverge

Takeaway: Returns may be bending, but if current trends hold, enterprises should expect continued gains in core model intelligence, industry-tuned execution and longer, more reliable agent runs over the next year, pushing more workflows from "assistive" to genuinely automatable.

3. What this week means for enterprises

3.1 “Best model” is now a workload question

Rough guide going into 2026:

Takeaway: A repeatable and efficient approach is needed to validate “which model is best for this job**”.

3.2 Cost-per-token is now strategic

Patterns that are stable enough to plan around:

For enterprises, the real metric is cost-per-successful-task: what you pay end-to-end to get work done to an acceptable standard. Tokens-per-watt and hardware efficiency matter for hyperscalers or those running their own clusters; everyone else feels them indirectly via pricing and capacity.

Takeaway: Model choice is about cost-engineering and capability-picking.

3.3 Architecture: Design for continual change

Given how fast the leaderboard rotates, three design principles to consider are:

  1. Abstract the model layer – put a gateway or internal abstraction in front of model calls so you can swap GPT-5.1 for Gemini 3 for Grok 4.1 for Sonnet 4.5 without rewriting half your stack.
  2. Route by workload – let coding agents favour Codex-Max/Sonnet 4.5; multimodal retrieval favour Gemini; low-stakes chat/marketing lean on Grok or open-weights. Don’t force everything through one endpoint.
  3. Version pin and test – pin model versions for critical workloads and maintain a small regression suite so you notice when post-training updates change behaviour.

Then align to your primary clouds:

A regulated Australian bank on Azure, for example, might end up with:

Takeaway: Architecture should enable teams to plug models in and out without impacting business logic.

4. A note on open-weights vs closed models: where they actually fit

Open-weights models (often labelled “open-source”) have had a big year. The standouts: DeepSeek-R1/V3, Llama 3.1, Kimi K2, Qwen3, OpenAI's openweights models and Gemma 3.

On capability:

They’re a great fit when:

They’re a more cautious fit when:

Takeaway: Use open-weights where control, sovereignty and cost win; closed models where peak capability and safety ecosystems win – and be explicit in your risk artefacts about jurisdiction and sovereignty risk for different model families.

5. Where this leaves you

If you’re running AI strategy, this week is a useful reset:

  1. Assume churn at the top. Gemini 3 Pro, Grok 4.1, GPT-5.1 and Claude Sonnet 4.5 can all claim “best” on some slice. That may not stabilise for some time.
  2. Architect for multi-model by default. Expect to mix 3–5 models across your estate and swap them over time. Align your API layer, security and observability with that reality.
  3. Treat pre-training and post-training as separate bets. Scale lifts the ceiling; post-training (RLHF, tools, safety, ORL) shapes behaviour, factuality and agent performance. When you pick vendors, you’re really choosing their post-training philosophy, not just their FLOPs.
  4. Be deliberate about open-weights and geopolitics. Open-weights are good enough to matter, and some of the strongest are Chinese. That doesn’t mean “don’t use them”, but it does mean conscious, documented decisions in regulated or government-adjacent sectors.
  5. Keep risk and finance close. Token prices, reasoning-effort settings, caching, time-horizon curves and power constraints are now economic variables, and regulators are increasingly interested in which models you use where, and why.

If it all feels like a lot, that’s normal for where we are in the cycle. The upside is that we’re still early. This week’s launches are one frame in what looks like what could be a decade-long reel of steady, compounding progress.

The trick isn’t to guess who “wins” the model war. It’s to build architecture, governance and skills that can ride the curve – whoever happens to be on top next quarter.