AI Weekly Recap: Four Major Vendors Cut Prices Together, Copilot's Paid Conversion Rate Just 4.5%

Introduction: The AI Industry Enters a Week Packed with Events

Over the past week, the AI industry has seen a string of major developments—OpenAI publicly released the GPT-5.6 family, xAI launched Grok 4.5, Google's Gemini 3.5 Pro was delayed, Microsoft's Copilot conversion data was exposed, and DeepSeek's legacy API is heading toward retirement. More than any single benchmark score, the most noteworthy signal of the week is this: near-frontier inference costs dropped simultaneously across four vendors, which will materially change which tasks are economically worth automating.

"Near-frontier inference" refers to inference services whose performance approaches that of today's most capable models but whose cost has fallen dramatically. There's a clear economic logic behind it: whether AI automation is worth deploying essentially comes down to the ratio between "the cost of a model call" and "the value the task generates." Unlike traditional software automation, LLM inference has a cost structure where every call incurs a marginal cost, and that cost is positively correlated with task complexity (token consumption). This means there's an "automation economics line"—only when the cost per call falls below a certain critical fraction of the cost of human processing does large-scale deployment become commercially viable.

Notably, this cost structure differs fundamentally from traditional SaaS licensing fees: traditional SaaS is a fixed subscription model where the marginal cost per call approaches zero; LLM inference, by contrast, is billed per token, meaning higher usage leads to higher cost. Before scaling deployment, enterprises must precisely calculate the token-consumption-to-output-value ratio for each business scenario. When the price per million tokens falls from tens of dollars to single digits, the automation ROI of tasks that previously required human labor—document review, code generation, customer service replies—undergoes a qualitative shift: scenarios that were previously below the threshold (such as processing tens of thousands of moderately complex documents per day) can suddenly cross the break-even point, triggering nonlinear growth in deployment demand. This is precisely why the industry views this week's collective price cuts as an "inflection point" rather than an ordinary promotion.

This article is based on the official blogs of OpenAI, xAI, and Meta, along with this week's coverage from media outlets including Euronews, Bloomberg, TechCrunch, and CNBC.

OpenAI's GPT-5.6 Family: Three New Models with Distinct Positioning

After concluding its limited partner preview on July 9, OpenAI officially released the GPT-5.6 family to the public, comprising three models with different positioning:

• Sol: aimed at frontier reasoning scenarios;
• Terra: delivering previous-generation flagship performance at roughly half the cost;
• Luna: focused on speed and low cost.

Beyond the text models, OpenAI also introduced GPT-Live-1, a voice model supporting full-duplex operation. Unlike traditional "half-duplex" voice AI—which must wait for the user to finish speaking before it begins processing, and can only respond after processing is complete—full-duplex models can handle audio input and output simultaneously, supporting real-time interruption, interjection, and emotional awareness, fundamentally transforming the human-machine conversation experience.

Full-duplex voice interaction requires solving several core technical challenges at the system level: echo cancellation (as the model receives input while outputting audio, its own voice gets picked up again by the microphone, forming an interference loop), interruption-handling logic (determining in real time whether a user's interjection is an "intentional interruption" or "background noise"), and a streaming inference architecture (beginning to generate a response before receiving complete user input).

From a system architecture standpoint, this challenge runs deeper than it appears: traditional voice AI uses a three-stage pipeline of "ASR (speech recognition) → LLM (language model) → TTS (speech synthesis)," with each stage processed serially, and the accumulated latency often exceeds 2 seconds. True full-duplex, however, requires integrating the audio encoder, language model, and audio decoder into a single end-to-end streaming system, so that "perception" and "generation" can run in parallel. Although GPT-4o previously supported a voice mode, its underlying flow was still a serial "listen → think → speak" process. This means human-machine interaction can be upgraded from a "voice version of text input" to a genuine "real-time conversational agent," which is significant for scenarios such as customer service, companionship, and real-time translation. Launched alongside it, gpt-realtime-2.1 reduces p95 latency (the response time for 95% of requests) by roughly 25%, further reducing the "stutter" that the vast majority of users perceive in real-world use.

Terra, positioned as "flagship performance at half the cost," is a key part of this week's wave of LLM price cuts.

xAI Grok 4.5: Joining the Price-Cut Race, Independent Evaluation Still Pending

xAI released Grok 4.5, reportedly co-trained with Cursor, priced at $2 per million input tokens and $6 per million output tokens. The company claims it reaches Opus-level performance on tasks such as programming, law, and finance.

However, some caution is warranted: there is currently no third-party independent evaluation data. Until results are borne out in practice, "matching top-tier models" is more of a vendor's marketing positioning, and its actual capabilities remain to be verified. But judging by the pricing, Grok 4.5 has clearly joined this round of price-cutting competition. The detail worth noting is "co-trained with Cursor"—this hints that xAI may have adopted a reinforcement learning fine-tuning strategy targeting specific tool-use scenarios, similar to DeepSeek's specialized training approach for mathematical reasoning, rather than a purely general capability improvement. This kind of training oriented toward specific workflows often produces performance beyond what the parameter scale would suggest in vertical-scenario evaluations, but carries uncertainty in cross-domain generalization.

Turmoil at Google: Gemini 3.5 Pro Delay and Talent Exodus

Google pushed back the release of Gemini 3.5 Pro to July 17, calling it a complete architectural rebuild that will support an ultra-long context window of 2 million tokens. Two million tokens means the model can process roughly 1.5 million English words in a single call—about ten times the length of the entire War and Peace—which is groundbreaking for scenarios such as full-text review of legal documents and analysis of very large codebases.

One of the core goals of the architectural rebuild is precisely to solve the "attention dilution" problem pervasive in ultra-long contexts. Transformer's self-attention mechanism can, in theory, attend to information at any position in a sequence, but in practice there is a significant positional bias—the model retains stronger memory of content located at the head and tail of the sequence, while its attention to middle passages drops markedly, a phenomenon Stanford researchers named the "lost in the middle" effect. For a 2-million-token ultra-long context, if you stuff an entire legal contract into the context, a key clause could be "overlooked" by the model simply because it sits in the middle of the document. This makes whether ultra-long context is genuinely usable (rather than just a numbers game on paper) the core test of whether this architectural rebuild succeeds.

More alarming is the turmoil at the talent level: within a single week, four senior DeepMind researchers departed one after another—Shazeer joined OpenAI, while Jumper, Adler, and Pritzel moved to Anthropic. Among them, Noam Shazeer was one of the co-authors of the landmark 2017 paper Attention Is All You Need. Understanding this paper's significance requires knowing what technology it replaced: the previously mainstream LSTM (Long Short-Term Memory network), while capable of processing sequential data, suffered from gradients gradually vanishing during backpropagation across time steps (the vanishing gradient problem), severely limiting the effective context length. The Transformer, through the self-attention mechanism, allows any two positions in a sequence to directly compute their relevance, fundamentally eliminating this limitation, while its highly parallelizable computation graph enables full utilization of GPU compute—it is precisely the combination of these two properties that laid the technical foundation for later ultra-large-scale training such as GPT-3 and PaLM. Shazeer's subsequent contributions in the field of sparse Mixture-of-Experts (MoE) models have also been widely adopted by mainstream models such as GPT-4 and Gemini. The reason the movement of top AI researchers is so sensitive is that they take with them not only their individual capabilities but also their judgment on frontier directions and their unpublished research ideas—these "metacognitive" capabilities are the truly irreplaceable core assets. Affected by this, Alphabet's market capitalization briefly evaporated by around $225 billion, reflecting the capital markets' strong endorsement of the judgment that "talent is the moat"—in the AI field, the flow of top researchers is often seen as a leading indicator of shifts in the technology competition landscape.

Microsoft Copilot: The Market Warning of a 4.5% Conversion Rate

Microsoft plans to consolidate its various Copilot applications into a unified product before August. But the most instructive figure this week is: of the 450 million M365 seats, less than 4.5% have converted to paid Copilot.

This number reveals the core contradiction of AI productization, and also reflects the deeper logic of enterprise software procurement: enterprise budget approval requires quantifiable ROI proof, yet the value of a general-purpose AI assistant is often a diffuse efficiency gain that is hard to reflect on financial statements. By comparison, GitHub Copilot's enterprise adoption rate in code-completion scenarios is significantly higher than that of general-purpose Copilot—precisely because the value of "reducing the time developers spend searching documentation" can be clearly quantified by engineering teams.

The tension between "being able to do everything" and "doing one thing well" is thus apparent: a general-purpose assistant's value proposition is "lowering the marginal cost of information processing," but the logic behind enterprise payment is "solving a specific, quantifiable business pain point." When a tool's selling point is "helping you write emails, summarize documents, and generate slides," users struggle to articulate the ROI during budget approval. By contrast, vertical AI tools focused on a specific workflow (such as contract review, code completion, or customer-service automation) find it far easier to complete the chain from "value perception → payment decision." This aligns with many practitioners' observations—the market's real demand lies in task-specific automation, not a general-purpose assistant that "helps a little with everything." For developers serving SMB customers, this judgment is especially worth taking seriously.

Meta and the Open-Source Ecosystem: Muse Image and Ollama's Funding

Meta released the first model from its Superintelligence Labs, Muse Image—an agent-capable image generation model that can call tools such as search and code and perform self-optimization. The model is trained by default on public Instagram photos (with an opt-out option), once again sparking discussion around data licensing and privacy.

The open-source space was equally active:

• Ollama completed a $65 million Series B round, with 8.9 million monthly active developers—a scale indicating that local inference has moved from a niche for geeks into the mainstream developer community, and Ollama is becoming the infrastructure layer of the local LLM ecosystem;
• Gemma 4, via multi-token prediction, runs about 90% faster in Ollama on Apple Silicon.

Multi-token prediction (MTP) is a training-paradigm improvement proposed by Meta: traditional autoregressive language models predict only the next token at a time, and at inference must execute N steps serially to generate N tokens. MTP introduces additional prediction heads during training, letting the model simultaneously learn to predict the next 2 to 4 tokens, thereby achieving a degree of parallelization at the inference stage. This idea shares a common thread with speculative decoding—the latter boosts throughput by having a small model pre-generate a draft and a large model verify it in parallel, while MTP internalizes the parallel prediction capability into the model's own training, reducing reliance on an extra draft model.

Apple's M-series chips' Unified Memory Architecture (UMA) solves another bottleneck: in traditional PC architecture, CPU memory and GPU memory are physically separate, and large-model weights must be frequently shuffled between the two—a bandwidth bottleneck that is especially pronounced during inference. UMA has the CPU and GPU share a single high-bandwidth memory pool, entirely eliminating data-transfer latency and making it feasible to run models with fewer than 70B parameters locally. The combination of these two technologies is precisely the underlying reason for the roughly 90% speedup.

The continued advances in local inference and on-device optimization are making open-source models increasingly competitive on both cost and performance.

Developer Alert: DeepSeek's Legacy API Retires on July 24

For developers using DeepSeek, an important reminder: deepseek-chat and deepseek-reasoner will be officially retired on July 24.

The company provides a one-line-of-code migration path, but there's a detail to note: deepseek-reasoner maps by default to the thinking mode of v4-flash, not v4-pro. Heavy reasoning workloads should not blindly trust the alias mapping and should explicitly switch to v4-pro to avoid a performance gap.

This is also an important reminder about engineering architecture—abstract your model layer. The idea of an "abstract model layer" stems from the Dependency Inversion Principle in software design: high-level business logic should not depend directly on low-level implementation details but should depend on abstract interfaces. In LLM application development, LiteLLM is one of the most mainstream open-source implementations today; it provides a unified OpenAI-compatible interface, and the backend can be seamlessly switched to dozens of providers such as Anthropic, Google, Cohere, and local Ollama, with zero changes to business code when switching. LangChain and LlamaIndex also offer similar capabilities.

Adopting this kind of abstraction layer brings another engineering advantage: against the industry backdrop of model iteration cycles now measured in months, it becomes convenient to A/B test different models and to incorporate performance evaluation and cost accounting into a routine engineering process. Model pluggability has shifted from a "nice to have" to a basic engineering requirement for production environments—DeepSeek's retirement event is a textbook warning of the risk posed by hardcoded model aliases: any production code that depends directly on a specific model name string harbors an uncontrollable single point of failure.

Analysis: The Significance of the Price-Cut Wave Far Exceeds Benchmark Scores

The most crucial judgment of this week is: the introduction of Terra and Grok 4.5 pricing while cutting costs matters far more than any single benchmark score.

When near-frontier inference costs drop simultaneously across four vendors within a single week, the boundary of "which tasks are economically worth automating" shifts accordingly. Scenarios that previously couldn't be scaled because of prohibitive costs may suddenly become viable. This change can be understood mathematically: when the per-token price drops by 50%, a scenario with an ROI of 0.8 (below the deployment threshold) may see its economic viability jump directly to 1.6, triggering a leap from "not worth doing" to "must do"—rather than a linear, gradual improvement.

On the other side, Microsoft's 4.5% conversion rate stands in stark contrast to the price-cut wave: both technical capability and cost are improving, yet general-purpose assistants still won't sell. The market's real signal is—providing deterministic automation value for specific tasks is the main storyline of today's AI adoption.

Key Takeaways

Key Takeaways