OpenAI vs. Anthropic: The Race for Customers Is Heating Up

Sam Altman Gets Nervous and Counters Anthropic

OpenAI is reportedly pushing out GPT-5.6 as early as today, July 7, with higher usage limits and stricter safeguards to lure away Anthropic users. Sam Altman is fueling expectations, comparing the model's mathematical progress to a child forming its first words. Nevertheless, Polymarket gives OpenAI only a 3 percent chance of having the leading LLM by the end of July; Anthropic holds the top spot with its Fable-5 models. In parallel, co-founder Greg Brockman outlines a future where agents handle tasks silently in the background, and users rarely have to click through menus. Brockman himself admits that the ChatGPT plugins failed in 2023 because the models were still too unreliable. And the systems still require heavy prompting, integrations, and human oversight, which is why OpenAI, Anthropic, and Microsoft maintain dedicated teams for enterprise adoption. → AI Breakfast

Synthszr Take: This 3 percent figure from Polymarket says more about the situation than any post from Altman. In early June, Anthropic had already surpassed OpenAI in valuation ($965 billion), and now it also holds the technical lead, with OpenAI responding with a rapid-fire point release. A quick drop with higher limits isn't a leadership position; it's a reaction to a loss of momentum. More interesting than the benchmark circus is Brockman's second message: When agents handle tasks in the background, the model itself becomes a commodity, and value shifts to orchestration. Therefore, anyone deploying AI today shouldn't bet on the next frontier benchmark score but should build their workflows so that changing a model is a matter of configuration, not a migration. That's the only position that will survive this race, regardless of who leads the synthzr charts in July.

Tencent Makes Hy3 Open Source, Beating GLM-5.2 Almost Everywhere

Tencent's Hunyuan team has released the full version of Hy3, a Mixture-of-Experts model with 295 billion parameters and 21 billion active per pass. The real news is in the license: Instead of the restrictive preview terms from April, the model is now available under Apache 2.0, without the previous exclusion of the EU, UK, and South Korea. In a blind test with 270 experts and 312 valid comparisons, Hy3 scored 2.67 out of 4, slightly ahead of GLM-5.1 (2.51), especially in frontend, CI/CD, and data work. However, Zhipu's GLM-5.2 retains the crown for agentic coding (SWE-bench Verified 84.2 vs. 78.0), which is hardly surprising given its roughly 744B parameters versus Hy3's 295B. In return, Hy3 leads the open-source field in agentic search (BrowseComp 84.2, DeepSearchQA 91.0) and tool orchestration. Tencent particularly emphasizes reliability: The hallucination rate dropped from 12.5 to 5.4 percent, and common sense errors fell from 25.4 to 12.7 percent. The model is available for free on OpenRouter for two weeks, with independent verification by Artificial Analysis still pending. → venturebeat.com

Synthszr Take: The license change beats any benchmark table. For a year, legal departments have been shelving the most powerful Chinese models before engineering teams could even finish their evals, because the terms excluded traffic from the EU, UK, and Korea. Apache 2.0 solves this exact problem, and for anyone serving European users, that's more important than three points on SWE-bench. The fact that GLM-5.2 remains ahead in repository coding while Hy3 wins in search and tool orchestration with half the compute per token illustrates this week's logic: You don't chain yourself to one model; you keep the intent layer clean and swap out the component as soon as a better one appears under a permissive license. The halved hallucination rate is the number that marks the path from a toy to a production tool. Anyone looking for a European failover option alongside Claude or GPT can test Hy3 against their own workflows during the two free weeks on OpenRouter and decide afterward. China's open-weight houses are now delivering production-grade models faster than most here can plan, and now they're delivering them in a legally sound way.

Mistral CEO Warns Against Closed AI Models

Arthur Mensch, founder of Mistral, warned against relying on closed AI models in a LinkedIn post. His argument: Those who sell proprietary models store more and more data, giving them a front-row seat to their customers' business processes. Some labs, according to Mensch, already have a habit of competing with their most successful customers using this very knowledge. He recommends keeping data in open systems, setting one's own access rules, and training one's own models. Support comes from Palantir CEO Alex Karp, whose manifesto coined the phrase: 'He who controls his weights, controls his destiny.' An experiment supports this thesis: Bridgewater and Mira Murati's Thinking Machines Lab fine-tuned the open-source model Qwen3-235B with their own evaluations and achieved 84.7 percent accuracy on financial documents, compared to 78.2 percent for the best frontier model, at nearly 14 times lower operating costs. The catch: Mistral, as the only relevant EU model, relies on this exact sovereignty narrative, even though about 30 percent of its shares are held by US investors. → Techpresso

Synthszr Take: Mensch is selling his own business model here, and yet he's right on the substance. For a hidden champion whose entire value lies in domain knowledge, the convenient connection to a closed model is the most dangerous temptation there is. You're not just outsourcing execution; you're giving the supplier insight into the very thing that makes you irreplaceable. Bridgewater's 84.7 percent isn't conclusive proof (both companies sell their own products), but it points in the right direction: internal expert knowledge that never ended up in the training data of large models beats the frontier model in a narrow domain. That's the raw material no competitor can buy, only earn. Those who organize their data in a machine-readable way and maintain clarity about their own 'why' in-house are making that decision now, not when the contract expires. You can buy the execution, but not the control over the weights.

GitHub Copilot Opens Up: Kimi K2 is the First Open-Weight Guest

GitHub Copilot is breaking its rule of only allowing closed models and is adding Kimi K2.7 as the first open-weight option to its program. The Superpowers plugin integrates the model directly into structured agentic workflows, allowing developers to use Kimi not just for coding but for entire task chains. Meanwhile, AlphaSignal shows how quickly the stack is changing: Sakana's multi-agent framework cracks 93 percent on Sudoku, where individual baselines get stuck at 11 percent. On top of that, there's a wave of cost-saving tricks, like pxpipe, which reduces Claude code bills by up to 70 percent by sending context as an image instead of text. OpenMed delivers 755 tokens per second on-device for redacting clinical data, and Mistral solves 587 out of 672 Putnam problems at a tenth of the cost. And DeepSeek follows up with R1, a freely available reasoning model on Hugging Face. The common thread: Open-weight is moving to the center of the most widely used dev tools. → AlphaSignal

Synthszr Take: When Kimi K2.5 was orchestrating agent swarms in January, it was still a side note for early adopters. Now, K2.7 is in Copilot, the tool used by millions of developers every day. This is the real tipping point, because distribution beats benchmarks. An open-weight model in the standard toolkit of a Microsoft subsidiary is a signal of legitimacy that the closed-model camp won't like, because it erodes the willingness to pay for proprietary endpoints right where it originates. Sakana's 93 versus 11 percent also shows where things are headed: many small agents beat the one big brain, and small models run locally and cheaply. Anyone planning their dev pipeline today can test Kimi in Copilot this week and directly compare the cost curve against a closed model. The question is no longer if open-weight will arrive, but how quickly closed-model providers will adjust their prices.

SiliconFlow: China's 'Token Factory' Goes Public, Backed by Alibaba and Huawei

On June 30, SiliconFlow (硅基流动) filed for an IPO in Hong Kong, just 35 months after its founding. The business model: no proprietary models, no proprietary apps, but the layer in between. Using its proprietary SiliconLLM engine, the company orchestrates chips from Nvidia, Huawei Ascend, Biren, MetaX, and Moore Threads, runs DeepSeek, Qwen, Kimi, and MiniMax on them, and sells the whole package by the token to developers and enterprises. Revenue in 2023 was 55.3 million yuan (up 653 percent), the number of paying customers shot up from 2,454 to 716,000, while the net loss swelled to 345 million yuan and the gross margin flipped from 39.4 to minus 24 percent. At a valuation of 7.7 billion yuan, this corresponds to a P/S multiple of 140, three to four times more expensive than its US counterparts Fireworks AI and Together AI. The list of shareholders reads like a map of the supply chain: Alibaba, Huawei Hubble, Meituan, and SenseTime are on board, with Alibaba and Huawei being both chip suppliers and direct competitors through their own MaaS offerings. At the end of the year, 172 million yuan in cash remained. → Hello China Tech

Synthszr Take: The number that explains everything isn't the 140x multiple, but the R&D-to-revenue ratio of 378 percent. For every yuan earned, SiliconFlow burns an additional 24 fen, and its public cloud division, which carries the whole 'token factory' story, runs at a gross margin of minus 119 percent. The problem runs deeper than the cost structure: without its own model, it lacks pricing power; without its own chip, it lacks cost leverage; and without a cloud ecosystem, there's nothing to cross-subsidize the free vouchers (54 million yuan to free users alone). Alibaba and Huawei collect the rent for the GPUs and compete for the same customers with their own MaaS offerings, so the purchased growth flows partly directly into competitors' pockets. The bet is that China's fragmented chip landscape needs a neutral intermediate layer, and SiliconFlow wants to occupy this position before Ascend or Biren have their own software stacks ready. As became clear during the token price war at the end of May, the open-source price competition squeezes the margins of these intermediaries from above, while chip manufacturers can pull the rug out from under them. Anyone buying in here is buying a narrow window of opportunity against 172 million yuan in remaining liquidity, and that window is closing faster than the P/S multiple admits.

Meta's 'Pocket': One Prompt, and the Metaverse Returns Through the Backdoor

Meta has released 'Pocket,' an app that builds playable, interactive mini-games from a single text prompt, with no coding required. This is made possible by the team behind 'Gizmo,' which Meta acquired earlier this year: The generated 'Gizmos' react to touch, device tilt, camera, and surroundings, making them easy to extend into AR and VR spaces later on. The app's social feed works on TikTok logic, except here users generate and remix instead of just consuming. After the failure of Horizon Worlds, Meta is clearly targeting the short attention span of Gen MZ, aiming for a second attempt at the metaverse. The real plan is the creator economy: create, go viral, and monetize through virtual items, sponsorships, and paid remixes, similar to the model of Roblox and Fortnite, but accelerated by GenAI. And the data from Pocket flows back to the Superintelligence Labs, which will use it to train the next models. → Trendium.ai

Synthszr Take: Meta has understood that the first metaverse failed for the wrong reason. People didn't want to enter empty VR rooms; they wanted to do something convenient and fun in five seconds. Welcome to the casual economy: the scarce resource is attention, not the digital good, and a prompt-to-game loop hits this nerve perfectly. Anyone who opens the app this week can build a playable Gizmo during a coffee break that would have previously required a studio and weeks of work. The catch is in the architecture: you create freely, but the traffic, transaction fees, and especially the training data all converge at Meta, just like with the two-billion-dollar Manus acquisition in December. The realistic counter-strategy is to use Pocket as a launchpad and deliberately build one's own brand across platforms, rather than getting locked into a closed system. The tenfold added value lies not in the prompt itself, but in the human context that no model provides: a story that only you can tell.

Design System Debt

The UX Collective Newsletter nails an observation that designers have been preaching for years: A design system is not a deliverable, but a product with its own owner, budget, and governance. The core of the latest issue revolves around Figma, which is now making this debt everyone's problem. Previously, a single person would quietly pay off the design system debt in files no one else opened. Now, the price is paid by the engineer whose generated component comes out wrong, the PM whose timeline breaks with a flawed handoff, and the marketer whose banners end up off-brand. Additionally, the issue provides an applicable framework with 39 principles for Human-AI Interaction and a widely-read text on re-evaluating software engineering work, which names implementation-heavy generalists as the big losers. → The UX Collective Newsletter

Synthszr Take: What's really happening here is a redistribution of costs, and it's suddenly hitting the people with the budget. The argument was always correct and almost always lost because those who held the purse strings never felt the failure firsthand. Now they feel it. When AI generates components in minutes, a bad design system multiplies errors in fast-forward instead of slowing them down. This is experience debt mechanics in its purest form: production costs decrease, maintenance costs remain, and the bill moves from the designer's corner to half the organization. Anyone who continues to treat the design system as a side issue is buying into a source of error that gets more expensive with every generated component. Funding the design system this week and giving it a real owner is the cheapest investment on the table right now.

Claude Ports 'Command & Conquer' to iPhone in Hours

Ammaar Reshi, Lead Product and Design at Google AI Studio, has ported the 2003 real-time strategy title 'Command & Conquer: Generals Zero Hour' to the iPhone and iPad. The game runs natively on ARM64, without any emulator, including the campaign, skirmish, and the 'Generals Challenge' with touch controls. The graphics pipeline translates DirectX 8 to Apple's Metal API through several intermediate steps. Reshi used Anthropic's Claude Code with Fable 5 for this; the first build was ready in about 40 minutes, followed by 'a few hours' of debugging. He used up his entire Claude Max quota in two days. He has published the full source code as open source on GitHub (you have to bring your own assets, which are available for about $5 on Steam). Remarkably, Reshi works at Google and still reaches for a competitor's tool, commenting that you can love the AI space and respect the competition while still staying focused on building the best product. → Techpresso

Synthszr Take: 40 minutes to the first build for a port that would have previously kept a small team busy for weeks. This is precisely the point often overlooked in the whole LLM Wars debate about model benchmarks: the real leverage isn't the next point on the LiveBench table, but what a single person can accomplish over a weekend. The fact that a Google employee uses Claude for the task says more about the market than any press release. Tool loyalty is dead; what matters is what gets the job done. And using up the max quota in two days soberly shows the flip side: this velocity costs compute, and anyone using it seriously needs compute discipline instead of a gut feeling for the budget. Anyone with a stalled legacy port in their drawer this week should just throw it at a coding agent setup before adding another ticket to the backlog.

When Claude Thinks About Itself: Anthropic Discovers 'J-lens'

Anthropic has published a 16-author study titled 'Verbalizable Representations Form a Global Workspace in Language Models,' which describes a surprising structure within Claude. Using a new mathematical technique called the Jacobian Lens (J-lens for short), the researchers looked inside the neural network and found a so-called 'J-space': a small, privileged zone where the model holds concepts it can report on and reason with, surrounded by a much larger ocean of automatic processing. The whole thing mirrors the Global Workspace Theory by cognitive scientist Bernard Baars, according to which the brain works like a theater where only a tiny spotlight becomes conscious. Claude's processing is divided into three regimes: an early 'sensory' zone, a middle 'workspace' band with abstract concepts, and a final 'motor' zone for the outputted word. In five tests, the researchers show that this space resembles features of conscious access in humans: if you swap the internal vector for 'spider' with 'ant,' the answer to the number of legs changes from 8 to 6; if you replace 'France' with 'China,' all downstream circuits correctly provide data for Beijing. Remarkably, the J-space accounts for only about 6 to 7 percent of the representational variance but is almost solely responsible for whether the model can report on a concept. And it wasn't built; it emerged on its own during training. → venturebeat.com

Synthszr Take: The most exciting part is in the practical footnote, not the consciousness debate. Anthropic says the J-lens is already changing how they monitor their systems for security risks, and in the 'workspace' band, the model internally flags prompt injections before they become visible. This is a tool that makes the invisible intermediate step readable: Claude thinks 'Mars' before it names the fourth planet, and you can now observe that. Anyone running agents in production now has a lever that goes beyond the usual chain-of-thought protocol because it shows the silent internal activations instead of just the written-out scratchpad. Whether 'consciousness' is emerging here is the wrong question for operations; the right question is whether this 6 to 7 percent can be reliably read and manipulated without the model noticing. The answer seems to be yes, and that's precisely what turns interpretability from an academic nice-to-have into a building block for guardrails. I'm curious to see how quickly this goes from a research artifact to standard monitoring, because anyone who can read their model's silent workbench has a real advantage when it comes to the question of whether you can trust the thing.