In a complex regulatory environment, model freedom ensures that your workflows don't stop
When Anthropic dropped Claude Fable 5 and Mythos 5 last week, the reaction was immediate. Fable 5 scored 80.3% on SWE-bench Pro and an astonishing 85% on OS-World Verified for computer use, placing it state-of-the-art across nearly every benchmark. Mythos — the same underlying model with safeguards lifted — was even more remarkable, the kind of unconstrained research capability that explains why Anthropic had been restricting access under Project Glasswing for two months before public launch.
Kilo had day-zero access to Fable 5 for all of our users, and devs were thrilled by its autonomous capabilities. It was especially popular in Kilo’s Code Reviewer.
Then, just as suddenly as they were launched, the models disappeared. Three days after launch, the US government issued an export control directive and Anthropic had to pull both models globally — paying enterprise customers included. Because Anthropic does not have a system to instantly differentiate and verify the citizenship of all its millions of users globally, they were forced to abruptly disable the models for *everyone, *including U.S. citizens and their own employees, to ensure full compliance. Just like that, the most capable coding AI ever made widely available was gone. Anthropic disagrees with the directive and is working to restore access, but in the meantime, production doesn’t .
And here’s the thing: the frontier didn’t disappear when Fable and Mythos went dark. Recent months have produced a genuinely competitive field of models that are cost-effective, enterprise-ready, and available right now. You don’t have to wait.
It’s time to broaden your model mix, and here are some strong contenders for consideration that can support end-to-end enterprise workflows.
GPT-5.5 #
Some people seem to have just missed the news on GPT-5.5. The noisier the news cycle, the more difficult it is for even the biggest names to be seen. But OpenAI’s latest frontier play is remarkably powerful—and more efficient than comparable SOTA models.
OpenAI’s GPT-5.5 launched at the end of April as the first fully retrained base model since GPT-4.5, and two months of production deployment have given the community a clearer picture of where it actually earns its price. The headline benchmark is 82.7% on Terminal-Bench 2.0, which OpenAI has argued — with some justification — matters more for real coding work than the saturated academic tests that dominated 2024 leaderboards. On KiloBench, GPT-5.5 ranks first among all tracked models with a 74.2% completion rate, and at $72.63 per attempt it’s also less expensive than comparable Opus and Gemini models.
The long-context reasoning jump is the most underrated number in the launch: MRCR v2 at 1M tokens went from 36.6% on GPT-5.4 to 74.0%, more than doubling, which matters whenever you’re feeding it an entire codebase or a long chain of prior context.
For enterprise workflows, GPT-5.5’s two-month head start on embedding means it’s available broadly across the tooling you’re probably already using. The pricing doubled from GPT-5.4 at $5/$30 per million input/output tokens, but OpenAI’s own analysis, corroborated by independent benchmarking from Artificial Analysis and our own KiloBench tests, suggests the effective cost increase is closer to 20% once token efficiency is accounted for. Our recent split analysis (before Fable was blocked) showed that with reasoning on high, GPT-5.5 and Claude Fable 5 actually have similar execution and overall coding abilities. And Kilo is built for this type of enterprise LLM functionality. If you’re looking for a model to drive the Kilo CLI, GPT-5.5 is it.
Nemotron 3 Ultra #
NVIDIA’s Nemotron 3 Ultra is the open-weight story of 2026 so far, and it’s currently free to use in Kilo. Jensen Huang introduced it at Computex in Taipei, and the framing was deliberate: this is NVIDIA’s declaration that open-weight models can operate at the frontier. The architecture is a hybrid Mamba-Transformer MoE — 550B total parameters, 55B active — and it delivers what Jensen called “frontier smart” performance at 5x higher throughput than other open models in its class. On PinchBench, the agentic benchmarking tool, Nemotron 3 Ultra holds the top spot among open models with a 91% median success rate. Along those same lines, Artificial Analysis places Ultra at an intelligence index score of 48 — highest among US-based open-weight models, and competitive with proprietary models several times its price.
The production case for Nemotron 3 Ultra isn’t just the price (though the current “free” access is obviously hard to beat for a model of this class). It’s the combination of NVIDIA’s hardware lineage, the 1M token context window, and the open-weight architecture that lets enterprise teams self-host without data leaving their infrastructure. For organizations that have been burned by the Fable 5 situation — where a government directive could abruptly end access to a hosted model — hardware sovereignty is suddenly a much more concrete concern.
As our Nemotron 3 Ultra post noted when the model launched: Nemotron 3 Super has been a daily driver for a lot of Kilo users, but the 3 Ultra is where planning and long-horizon task quality finally close the gap with the closed frontier.
The model’s high instruction-following scores mean it’s reliable enough for structured, client-facing output, and the Nemotron Coalition, which we discussed on the recent Kilo webinar with NVIDIA, is helping to extend those capabilities to a wide range of industries and use cases.
MiniMax M3 #
MiniMax M3 launched earlier this month as well, and it immediately became the most interesting cost-efficiency story in the field. The headline: 59.0% on SWE-Bench Pro and 66.0% on Terminal Bench 2.1, edging past GPT-5.5 on both, at an API price of $0.30/$1.20 per million input/output tokens — approximately 1/40th the cost of Fable 5 at list pricing. The architectural innovation enabling this is MiniMax Sparse Attention (MSA), which cuts per-token compute at 1M context to roughly 1/20th of the prior generation, with more than 9x faster prefill.
In addition, MiniMax announced recently that the 50% discount at launch will now be permanent. So MiniMax M3 is priced with that “permanent discount” at just $0.30 per million input tokens. That’s 1/16th the cost of Opus 4.8.
To note: M3’s API is hosted by a Chinese lab, and as with any non-self-hosted model, data governance questions apply. But MiniMax was quick to highlight M3’s open weights after the Fable 5 takedown, and that’s the answer for teams where data residency is a hard requirement — the weights are available for self-hosted deployment, which removes the API dependency entirely and also means that a lot of providers can now host the model. Providers for M3 in the Kilo Gateway already include Fireworks AI, Morph and Parasail.
MiniMax M3 has been the third most popular Kilo model on OpenRouter since the Fable 5 news broke. And that’s in addition to the huge amount of tokens being used via MiniMax token plans, which you can now purchase with your Kilo Credits.
Kimi K2.7 Code #
Moonshot AI’s K2.7 Code arrived on June 12th — the same week the Fable situation unfolded — and it’s built for exactly the kind of production coding workflow Fable was meant to serve. The architecture is a 1-trillion-parameter Mixture-of-Experts model with 32B active parameters and a 256K context window, released open-weight under a Modified MIT license. The headline improvement over K2.6 is a 30% reduction in reasoning-token usage, which matters enormously at scale: every agentic loop that previously burned 1,000 tokens to think through a code change now burns ~700.
The MCP tool-use story is particularly compelling for Kilo users. K2.7 Code scored 81.1 on MCP Mark Verified, which evaluates correct tool invocation through the Model Context Protocol (MCP). That means CI checks, ticket updates, and multi-file edits in one loop, reliably. And at just $0.75/$3.50 per million input/output tokens, Kimi K2.7 is priced to be used every day.
Like MiniMax, Moonshot AI is a Chinese lab, but note that the Kilo Gateway supports a number of providers for all of the models we’ve discussed in the post, including for Kimi. Make sure to check this in your org settings to fit your internal guidelines.
The Fable 5 situation is a reminder that “most capable model available” is not the same thing as “most capable model you can rely on in production.” Reliability isn’t a fallback plan, it’s a feature.
This Ain’t Our First Rodeo
Remember *The Magnificent Seven, *the 1960 western featuring Steve McQueen as the “drifter and sharpshooter” Vin? In a moment of self-reflection, Vin pines:
The Old Man was right. Only the farmers win. We lose. We’ll always lose.
When Anthropic is dominating the news cycle, other labs and models can feel like wandering gunslingers without a home, trying to lasso the latest big target. But it doesn’t have to be that way.
The frontier is wider than one lab. With over 500 models available in the Kilo Gateway, plus extensive BYOK support, it’s never been easier to use the top models in a changing landscape.
The frontier is bigger than you might think. We have the models, and the done-for-you infrastructure, to prove it.