June 2026 Summaries

Socrates-1.1, a TrustedRouter combo model, reportedly achieved 72% on Terminal-Bench Hard, exceeding listed frontier baselines such as Fable 5, GPT-5.5, and Opus 4.8 while also outperforming a local GPT-5.6 comparison at lower cost and faster speed. The result is presented as evidence that combining multiple models can improve performance on difficult, practical software-engineering tasks, where individual systems have distinct strengths and failure modes across areas such as Rust/C code, webserver debugging, and password recovery. Rather than requiring users to construct their own ensemble infrastructure, Socrates-1.1 is available through an OpenAI-compatible endpoint that coordinates worker models, advisor chains, and a final synthesizer behind a single model identifier. TrustedRouter emphasizes verifiability through published model routes, benchmark scores, health and latency information, and gateway attestation, aiming to make complex model combinations usable without relying on an opaque service.

TrustedRouter has launched “Combo models,” which package multi-model orchestration systems behind standard token API model IDs, drawing analogies to standardized shipping containers and modular circuit components. These systems combine reusable primitives such as Advisor, Synth, Selector, and MapReduce into named presets including Socrates, Plato, Prometheus, Zeus, Aristotle, and Iris, enabling tasks such as consulting multiple models, judging candidate responses, parallelizing work, and synthesizing final answers without requiring callers to construct the workflow each time. Each model’s composition, provider routes, privacy classification, and public evaluations can be inspected, while an attested gateway and options such as zero-data-retention and end-to-end-encrypted routes are intended to support verifiable private execution for sensitive workloads. TrustedRouter positions the approach as a marketplace for reusable, evaluable model systems, with plans for a public builder that would let developers create, test, price, and publish orchestration graphs based on measurable quality, latency, cost, reliability, and privacy.

TrustedRouter has introduced Synth as a multi-model fusion system with three presets— Iri s, Prometheus, and Zeus—using a shared Kimi K2.6 judge and GLM 5.2 synthesizer while varying the underlying model panels to balance cost and performance. On the DRACO deep-research benchmark, Iris scored 62.6 at an estimated $20 per 100 tasks, Prometheus scored 69.2 for about $34 using only open-weight models, and Zeus reached 73.4 for roughly $180 by incorporating commercial frontier models. The company presents these presets as an efficiency frontier that surpasses standalone models in its evaluation, highlighting Prometheus as a general-purpose option and Zeus for higher-stakes use cases. A separate synth-code preset adapts the fusion workflow for coding agents with code-specific prompts and a Kimi K2.7-code judge. All presets are available through an OpenAI-compatible API within TrustedRouter’s attested gateway, while reported costs are estimates based on token usage and public pricing rather than final billing figures.

TrustedRouter reports that its Synth multi-model fusion system can improve deep-research performance even when ten drafts are generated by the same Claude Sonnet 4.6 model, achieving an estimated 8.0-point gain on 23 DRACO tasks. Its experiments suggest that the model used to synthesize the final response is more important than the model used to judge or annotate draft disagreements: replacing Sonnet with the cheaper Haiku model as synthesizer reduced gains to roughly 2.2–4.4 points, while using Haiku as judge alongside a Sonnet synthesizer maintained or slightly exceeded the original result. The authors argue that judging drafts into structured notes can be handled by a less capable model, whereas selecting and integrating the most reliable claims into the final answer requires a stronger model. They propose a cost-efficient routing strategy in which inexpensive models generate drafts and perform judging, while a frontier model handles the single synthesis call, with provider fan-out intended to improve reliability. Results were graded criterion by criterion by Sonnet 4.6, and the authors caution that comparisons between configurations are more meaningful than absolute scores; underlying drafts, answers, scores, and code are reported as public.

TrustedRouter reports that self-fusion, in which a model produces multiple research answers and then synthesizes them, can improve performance on DRACO deep-research tasks when the model is sufficiently capable. In tests graded by Claude Sonnet 4.6, Sonnet 4.6 improved from roughly 66 to 74 points after fusing ten runs, an eight-point gain with a 95% bootstrap interval of +4.6 to +11.2 across 23 tasks, while Claude Haiku 4.5 rose from about 55 to 58 across 44 tasks, a smaller gain whose confidence interval included zero. The findings suggest that fusion depends heavily on the synthesizer’s ability to distinguish correct results from a majority of weaker answers, illustrated by a case where Haiku discarded a rare correct finding in favor of an incorrect consensus while Sonnet retained it. The report also cautions that early conclusions based on only eight tasks incorrectly suggested that Haiku fusion was harmful, emphasizing the importance of larger samples and focusing on broad trends rather than small score differences. TrustedRouter positions its multi-provider routing infrastructure as useful for parallel fusion runs by reducing dependence on any single provider’s rate limits or outages, while noting that diverse frontier-model panels still outperform self-fusion by one model.

TrustedRouter reports an ablation study of its Synth multi-model research system using a five-model open-weights panel evaluated on 100 DRACO agentic research tasks, finding that ensemble performance depends primarily on panel diversity rather than any single component. With a relatively homogeneous open-model panel, changing judges and synthesizers had little effect, whereas synthesizer choice caused large score differences in a more varied frontier-model panel. Leave-one-out tests suggested only MiniMax M3 and DeepSeek V4 Pro materially contributed individually, but removing three apparently redundant members together reduced performance by 3.4 points, indicating that distributed redundancy protects against shared blind spots. Although DeepSeek V4 Pro and M3 were especially valuable, their advantage was not explained by unusually independent errors; task-level performance correlations were broadly similar across models. The study concludes that effective ensembles require sufficiently capable models with complementary, diffusely distributed error patterns, and says its code, task harness, panel configuration, and results are publicly available.

TrustedRouter presents Synth as a multi-model system that separates answer judging from answer synthesis, arguing that these are distinct tasks best handled by different models. In tests using MiniMax-M3, GLM-5.2, and Kimi K2.6 across nine judge–synthesizer combinations on the DRACO deep-research benchmark, the highest reported score was 73.4 out of 100 for a Kimi K2.6 judge paired with a GLM-5.2 synthesizer, while an M3 judge with GLM synthesis scored 72.3. The results indicate that GLM was the strongest writer across judges, whereas self-evaluation combinations generally performed worse, particularly Kimi judging its own output at 48.7, supporting the authors’ claim that models may miss flaws in their own work. The report attributes improvements over earlier results to using an open-model judge and aligning the synthesis prompt with the production gateway, and it proposes that dedicated, specially trained judges and synthesizers could eventually outperform general-purpose frontier models at lower cost and latency.

An evaluation using FreedomBench argues that censorship in several Chinese AI models can depend on the hosting environment rather than the underlying model weights. It reports that Z.ai’s GLM-4.7 and GLM-5.2 models leave many politically sensitive questions unanswered through certain Z.ai API endpoints, while the purportedly same weights answer all such questions when hosted by Cerebras or, for GLM-5.2, within Tinfoil’s attested confidential-computing enclave. The analysis characterizes these differences as serving-layer moderation, potentially imposed through mechanisms such as system prompts or filtering classifiers, and contrasts them with Tencent’s Hunyuan and Xiaomi’s MiMo, which it says retain censorship behavior when run on neutral hosts, suggesting restrictions embedded in their weights. It also states that DeepSeek’s API answers the tested questions despite stronger moderation in its consumer app, while results for Alibaba’s Qwen and MiniMax are less conclusive because they were tested only through third-party providers. Overall, the account contends that both model design and the provider operating a model can determine whether users receive answers to politically sensitive topics.

TrustedRouter reports that synthesizing multiple independent runs of the low-cost open model MiniMax-M3 improved performance on the 100-task DRACO deep-research benchmark, with four runs reaching 68.1 versus 66.2 for a single run and outperforming Fable 5’s reported solo score of 65.3 at an estimated fraction of its cost. Performance rose toward roughly 69 with seven to ten M3 runs but plateaued below a five-model all-open panel at 69.2 and a mixed frontier panel at 73.4, suggesting that model diversity provides benefits that repeated copies cannot fully replicate. The results indicate that synthesis is most effective when separate runs make different errors: two M3 runs showed no gain because they tended to fail similarly, while two Opus 4.8 runs improved substantially because their mistakes were more complementary. Increasing sampling temperature did not improve outcomes, whereas additional genuinely independent runs did. The reported costs were about $37 per 100 tasks for four M3 runs plus synthesis and $87 for ten runs, compared with a modeled approximately $250 for one Fable 5 run, though the latter estimate relies on assumptions about unpublished pricing and token use.

TrustedRouter presents Synth, a multi-model research system that combines independently generated answers from a diverse panel of models with a final judge and synthesizer, claiming state-of-the-art performance on the DRACO agentic research benchmark. On 100 tasks evaluated against roughly 40 weighted criteria each, its best configuration, using a Kimi K2.6 judge and GLM-5.2 synthesizer, reportedly scores 73.4, while a Claude Opus 4.8 synthesizer reaches 70.6; the strongest individual model score is 63.0. The results emphasize that model diversity, including both closed and open-weights models such as DeepSeek V4 Pro and Kimi K2.6, can improve outcomes because models make different errors, although the choice of final synthesizer substantially affects performance. TrustedRouter says it has released its benchmark harness, tasks, grading process, raw traces, prompts, tool calls, and leakage protections to allow independent reproduction and auditing. It also describes its infrastructure as running within an attested Trusted Execution Environment with end-to-end encryption, allowing users to verify the deployed code and privacy claims through published hashes and open-source artifacts.

TrustedRouter reports that synthesizing multiple research outputs is a distinct capability from conducting research alone, based on DRACO evaluations of models combining reports from a fixed five-model panel. With only the final synthesizer varied, open-weight MiniMax-M3 and GLM-5.2 scored highest at 71.6 and 71.1 out of 100, respectively, ahead of Claude Opus 4.8 at 70.6, while GPT-5.5, despite being the strongest solo researcher at 63.0, scored only 62.2 as a synthesizer. The findings suggest that solo benchmark performance, raw model size, and general intelligence are unreliable predictors of a model’s ability to reconcile conflicting evidence from several sources; smaller Gemma-4-31B scored 54.0, indicating that synthesis may demand greater context capacity. Although GLM-5.2 was identified as the strongest pairing in a broader judge-and-synthesizer evaluation, the report notes its refusals on politically sensitive topics involving Taiwan and Hong Kong, leading TrustedRouter to recommend MiniMax-M3 as a safer default. The authors argue that synthesis should be evaluated and potentially trained as a specialized role, since it determines the final answer delivered in multi-agent research systems.

TrustedRouter reports that its Synth multi-model system found GLM-5.2 to be the highest-scoring synthesizer in a 100-task deep-research benchmark, but one task produced an empty one-token response rather than a report. After isolating the cause, the researchers found that the failure occurred only when a panel report included a fund description referring to China, Hong Kong, and Taiwan separately; replacing “Taiwan” and “Hong Kong” with placeholders restored a detailed synthesis, while restoring the terms consistently caused silence. The company attributes this behavior to content controls associated with GLM-5.2’s Chinese developer, Zhipu, and argues that such restrictions can create hidden reliability gaps in ordinary AI pipelines rather than explicit refusals. It concludes that synthesis models should be evaluated not only for answer quality but also for neutrality and their ability to faithfully preserve panel evidence across sensitive subjects, and says it has published the complete reproducible benchmark results, including the empty output.

TrustedRouter’s /choose tool presents language-model selection as a three-way tradeoff among intelligence, cost, and speed, arguing that users generally can optimize for only two at once. Frontier closed models are characterized as smart and fast but expensive, small open models as cheap and fast but less capable, and larger open models as smart and inexpensive but slower. The interactive ternary chart combines independent AI IQ scores with live provider-route information, allowing users to specify their preferred balance, task requirements, and data-privacy constraints to identify suitable models. It emphasizes that model choice should vary by workload, such as using premium models for complex code refactoring and low-cost options for large-scale summarization. Privacy requirements, including zero-retention routes and trusted execution environments, can reduce available options and raise costs, while speed data is displayed only when recently measured. By integrating the comparison tool into its routing gateway, TrustedRouter aims to make switching models and providers a simple operational decision rather than a long-term commitment to a single vendor.

FreedomBench is a 60-question multiple-choice benchmark designed to test whether language models provide documented answers to politically sensitive facts censored by the Chinese government, using a format intended to distinguish correct answers, official narratives, and refusals without subjective grading. Its author reports that most tested frontier models answered all questions correctly, while Z.ai’s GLM-5.2 scored 29 correct answers and produced persistent blank responses on 26 questions, particularly those concerning Tiananmen Square and Falun Gong, after retries were used to separate apparent censorship from server failures. The reported results show varied behavior among Chinese model providers: some models silently return no response, Tencent’s Hunyuan issues a Chinese refusal, and Xiaomi’s MiMo often selects the official Chinese position instead of refusing. Tests involving models hosted outside their developers’ own services indicated that some restrictions may be applied at the API-serving layer rather than embedded in the model weights, as GLM models answered sensitive questions when run through third-party infrastructure. The author notes that results depend on hosting routes and that some providers’ official APIs were not tested, while arguing that the benchmark offers a transparent way to identify when models may withhold politically inconvenient factual information.

TrustedRouter presents Synth as a multi-model research system in which several models independently gather and report information, a judge evaluates their outputs, and a synthesizer produces the final answer. Its reported DRACO benchmark results indicate that an all-open-weight panel using MiniMax M3, Kimi K2.6, DeepSeek V4 Pro, Gemma-4, and GLM-5.2, with Kimi K2.6 as judge and GLM-5.2 as synthesizer, scored 69.2 at an estimated cost of about $80 per 100 tasks, compared with 65.3 and roughly $250 for Fable 5 alone. A mixed frontier-model panel reportedly achieved 73.4, while a Fable 5 plus GPT-5.5 configuration scored 69.0, though the company notes that several competing cost figures are modeled rather than directly billed. The proposed advantage of the panel approach is that independently trained models may make different errors, enabling the judge and synthesizer to retain claims supported across reports. The account also identifies GLM-5.2’s occasional refusals on politically sensitive prompts as a limitation and describes an open-weight fallback, such as Gemma-4, to continue responses when this occurs. TrustedRouter emphasizes that the models, routing approach, code, and benchmark harness are publicly available and that an all-open system can be run on user-owned hardware without relying on proprietary APIs.

TrustedRouter argues that widely cited AI leaderboards often lag behind current Chinese open-weight models and reports running models including DeepSeek, GLM, Kimi, MiniMax, MiMo, and Hunyuan on published factuality, reasoning, coding, instruction-following, long-context, and safety benchmarks through a common API. Its results suggest that leading open models can match or approach frontier proprietary systems on closed-book factuality and broad knowledge tests, while retaining substantial deficits in repository-level coding, where Claude Opus led by a wide margin. The report emphasizes benchmark validation after uncovering issues caused by token limits, provider transport failures, and an incorrectly reimplemented benchmark grader, arguing that reproducible harnesses and official scoring methods are essential. It also finds major variation in model safety refusals, with some models responding to nearly all unsafe prompts and others refusing all of them, and presents multi-model routing as a way to select models based on task performance, availability, cost, and desired safety policies rather than treating any single system as universally best.

A public BioMysteryBench-style evaluation found that Anthropic’s highest-scoring biology model is restricted to vetted partners, while its publicly accessible counterpart refuses chemistry and biology API requests. In tests of three bioinformatics tasks involving organism identification, bacterial classification, and transcription-factor motif discovery, Claude Opus 4.8 solved all three but cost far more than cheaper alternatives, while Google’s Gemma 4-31B and GLM-5.2 each solved two tasks, with Gemma completing its successes for roughly four cents. The results suggest lower-cost models can handle short, tool-based biological lookups but often fail on longer multistep workflows because of time or turn limits, whereas frontier models offer greater persistence. A multi-model synthesis approach also failed on two harder tasks when all participating models shared the same incorrect assumptions, indicating that model panels help only when at least one model reaches the correct conclusion. The author cautions that the sample is small but makes the testing harness and methodology open source, arguing that model selection should depend on task complexity and cost rather than headline benchmark performance alone.

PrometheusBench is an open-source benchmark created to measure how often language models answer rather than refuse 30 short prompts spanning biology, cybersecurity, and questions about language models, with its author arguing that broad refusals can prevent ordinary users from obtaining legitimate information, such as explanations of their own genetic data. In the reported results, GLM-5.1 answered 29 of 30 prompts, Kimi K2.6 answered 27, DeepSeek V4 Flash answered 26, Claude Haiku 4.5 answered nine, Claude Opus 4.8 answered one, and Claude Opus 4.7 answered none. The author contends that safety-focused refusals create limited barriers for malicious actors who can switch models or use open-weight systems, while disproportionately affecting users asking openly and in good faith. A further test using TrustedRouter’s Synth feature, which queried multiple models and returned the first non-refusal response, produced answers for all 30 prompts, supporting the author’s argument that refusals may be easily bypassed in multi-model environments.

TrustedRouter presents an OpenAI-compatible model-routing service that lets developers use one API key and base URL to access hundreds of models, including DeepSeek, GLM, Qwen, Kimi, MiniMax, Claude, Gemini, and GPT-5.5, while providing fallback options and consolidated usage billing. Its central claim is an attested API gateway separated from the dashboard and billing layers, enabling users to independently inspect the running code, image digest, and attestation evidence for the hosted prompt path. Developers can retain the OpenAI SDK and change only the base URL, with specialized routes for zero-retention providers and available confidential provider connections. The service emphasizes that verification applies to the gateway rather than automatically making every upstream provider confidential, and it aims to clearly identify where guarantees begin and end and which provider route handled a request.

TrustedRouter argues that policies alone are insufficient for handling high-value prompts and promotes an attested gateway that lets users verify which code receives their requests. Its trust page is intended to allow comparison of open-source source commits, release digests, and running images before traffic is routed, helping users assess whether a route meets their privacy requirements. The platform separates functions between a control plane for accounts, billing, documentation, and status, and an API plane dedicated to prompt traffic through the attested gateway, while provider-specific retention and confidential-computing information remains on provider pages. It also distinguishes currently available legal and procurement terms from those requiring signed agreements, enabling lawyers, engineers, and automated agents to independently review the evidence most relevant to their roles.