Grok 4.5 Launch: Coding AI Market Reshuffle, Anthropic Claims Half the Territory

Grok 4.5 Arrives: A Dual Evolution in Coding and Agents

The AI industry is entering an intense wave of releases. Grok 4.5 has officially launched, focusing on dedicated coding and agent capabilities. Most notable is its extreme cost control—each task costs just $0.49, roughly 90% cheaper than comparable products.

Musk claims Grok 4.5 delivers performance on par with competitors, but at greater speed. The model's training strategy is quite clever: it was trained on real interaction data from Cursor's engineers, halving token usage and bringing overall costs to less than half of competing products. This "learning from real development scenarios" approach is essentially a data flywheel strategy. A data flywheel is a model iteration mechanism that continuously self-reinforces through user behavior data: users generate high-quality behavioral data by using the product, this data is used via knowledge distillation to train leaner models, improved model performance attracts more users, which in turn generates more training signals—forming a virtuous cycle. Knowledge distillation, proposed by Hinton et al. in 2015, works by having a smaller model learn the output probability distribution of a larger model rather than hard labels, preserving the model's "knowledge" while compressing parameter count. As a professional coding tool with a large number of daily active developers, Cursor's user interaction data is highly specialized and authentic, making the signal quality for distillation training far superior to that of generic web-crawled data—this is precisely why Grok 4.5 can maintain performance while halving token usage, and it is becoming the new paradigm for iterating AI coding models.

Notably, the data flywheel effect is especially pronounced in vertical domains. General internet-crawled data, while massive in volume, is noisy and semantically ambiguous, making it hard to reflect real engineering decision logic. Interaction data from professional IDEs, on the other hand, naturally contains the complete feedback loop of "problem posed → code generated → user accepts/rejects/modifies," where every acceptance or modification is a high-quality implicit preference label. This data quality advantage, combined with the parameter compression benefits of knowledge distillation, enables Grok 4.5 to achieve coding reasoning capabilities approaching those of large models at a smaller scale, thereby driving inference compute costs to the lowest level in the industry. This approach offers important methodological value for resource-constrained AI teams: rather than blindly stacking parameters, it is better to cultivate high-quality domain-specific data.

Meanwhile, OpenAI's GPT 5.6 series launched simultaneously, forming a clear product-line tiering: the Sol version offers stronger performance at the same price, the Terra version delivers comparable performance at half the price, and Luna pushes further downmarket. This "multiple tiers within a generation" pricing strategy reflects the fine-grained operational thinking of leading vendors toward different user segments under cost pressure. It's worth noting that this tiered pricing is essentially a direct mapping of the inference compute cost curve—as inference acceleration technologies like quantization compression and speculative decoding mature, the same generation of base model can spawn product variants with varying performance and cost through different precision configurations, allowing model providers to cover the full spectrum of users from individual developers to large enterprises, rather than relying on a single flagship product.

Coding Market Reshuffle: Anthropic's Powerful Rise

The most noteworthy development this round is the dramatic shift in the AI coding tools market. Cursor, which once held a 41% share, has seen its market share fall to 26%; meanwhile, Anthropic has surged with Claude Code, now capturing roughly 50% of the AI coding market. Claude Code's code contributions even exceed 7% of GitHub's total commits.

Claude Code's ability to capture such a large market share in a short time is closely tied to Anthropic's specialized optimizations in context window design and code reasoning pipelines. Unlike general-purpose conversational models, coding scenarios demand extremely high standards for long-context understanding, multi-file dependency tracking, and incremental editing precision. Anthropic's Constitutional AI training framework allows Claude to reject harmful code generation while maintaining a low false-rejection rate, creating a differentiated advantage in enterprise security and compliance scenarios.

Constitutional AI is a distinctive alignment method proposed by Anthropic in 2022. Its core idea is to have the model self-critique and revise its own outputs based on a set of explicit "constitutional principles," rather than relying entirely on human-labeled preference data. This method significantly reduces dependence on large-scale human annotation while achieving fine-grained safety boundary control in code generation scenarios—the model can distinguish between "commented-out example vulnerability code" and "genuine malicious code generation requests," keeping the false-rejection rate at a level that doesn't harm the development experience. This is especially critical for enterprise development scenarios that frequently handle security testing code and penetration testing scripts. From a competitive standpoint, Cursor's declining market share is not simply due to product quality degradation; the deeper reason is that Claude Code adopts a CLI-first interaction mode that better aligns with professional developers' habits, along with clear advantages in global comprehension of large codebases—two points that directly strike at the inherent limitations of the IDE-plugin form factor in complex engineering scenarios.

Particularly noteworthy: GitHub currently processes about 3 million code commits per day, and Claude Code contributing over 7% of these means it participates in generating or assisting more than 200,000 real code commits daily—AI-assisted programming has upgraded from an "efficiency tool" to "production infrastructure," completing a qualitative transformation.

Behind this lies Anthropic's astonishing efficiency leap. Improvements in AI inference efficiency typically come from multiple dimensions: model quantization (compressing weight precision from FP32 to INT8 or even INT4), KV Cache optimization, improved batch scheduling algorithms, and co-design with dedicated inference chips. Anthropic tripled its inference efficiency within nine months, very likely by combining multiple techniques above along with proprietary inference kernel optimizations, ultimately tripling the revenue generated per megawatt of compute and pushing editing profit margins close to 100%—far exceeding OpenAI. The company expects profits to surpass $1 billion in the third quarter of next year, with API revenue accounting for 80% of that.

This data reveals a key trend: in the AI application deployment phase, whoever can drive inference costs to the lowest will survive the price war. Anthropic profits through its API, while vendors relying on subscription models that need to subsidize free users face continuous capital burn pressure.

New Thinking on AI Self-Evolution

Industry experts propose that AI self-evolution should start from the external system harness rather than directly modifying model weights. To understand this judgment, one must first grasp the fundamental difference between the two approaches.

RLHF (Reinforcement Learning from Human Feedback) is currently one of the mainstream large-model alignment techniques: collect human preference annotation data → train a reward model → use the PPO algorithm to update language model weights with reward signals. This process is computationally expensive, time-consuming, and difficult to roll back once bias is introduced. The so-called "external harness" approach is entirely different—without altering model parameters, it uses external scaffolding systems (such as tool use, persistent memory modules, and automatic feedback loops) to enable the model to continuously improve its behavioral strategy during the inference phase. This is analogous to humans improving efficiency by establishing external workflows and SOPs rather than retraining their nervous system. Its engineering advantage lies in the ability to complete strategy iterations at hourly granularity, with each change being independently rollback-able, greatly reducing deployment risk in production environments—far more suited to real-world engineering conditions than "brute-force weight modification." DeepSeek's technical team also agrees with this judgment, offering the industry a technical route that differs from "stacking parameters."

From a software engineering perspective, the external harness approach aligns closely with the design philosophy of microservices architecture: breaking system capabilities into independently deployable, independently rollback-able modules, and combining them through standard interfaces to achieve complex behavior, rather than coupling all logic into a single monolithic system. For enterprise-grade AI deployment, this means being able to achieve rapid customization for specific business scenarios by adjusting tool-calling strategies, memory retrieval algorithms, and feedback evaluation criteria—all without touching the underlying model. This is why an increasing number of enterprises are choosing to build private agent systems based on open APIs, rather than waiting for model vendors to complete the next round of fine-tuning.

Funding and Compute: New Directions for Capital

The capital markets are equally lively. SambaNova completed a $1 billion Series F round at an $11 billion valuation, with its valuation increasing sevenfold in half a year. JPMorgan has signed on to use its inference infrastructure, and the SN50 chip will ship in the second half of the year. SambaNova's core competitiveness lies in its proprietary dataflow architecture chips—unlike the von Neumann architecture of GPUs, traditional GPUs require data to be shuttled repeatedly between memory and compute units, which easily triggers the "memory wall" bottleneck in large-scale matrix operations, causing compute to spin idle waiting for data. The dataflow architecture fundamentally restructures this logic: the compute graph is statically mapped to hardware circuits at compile time, and data flows along fixed paths between compute nodes, almost eliminating redundant data movement overhead. This concept originated from dataflow computer research in the 1980s and only became truly commercializable with the widespread adoption of advanced sub-7nm processes. Its energy efficiency in Transformer inference scenarios is reportedly 3-5 times that of same-generation GPUs, which is also the core technical barrier that won it large contracts among inference-intensive financial institutions like JPMorgan. SambaNova now serves over 360 leading enterprises including TSMC, with revenue doubling consecutively.

The energy efficiency gap between dataflow architecture and GPUs is especially pronounced in inference scenarios. GPUs were originally designed for graphics rendering, and their SIMD (Single Instruction, Multiple Data) parallel mode excels at large-batch matrix multiplication during training. But in online inference scenarios, requests often arrive in small batches or even one at a time, leaving GPU parallel resources heavily idle, while the data-movement overhead between HBM memory and SRAM is hard to compress, causing actual energy efficiency to fall far below peak specifications. SambaNova's dataflow architecture statically compiles the compute graph to hardware, directly connecting each compute node's input/output to adjacent nodes so that data need not transit through global memory, fundamentally eliminating this "movement tax." This advantage is especially evident in the alternating computation of the Transformer's attention mechanism and feed-forward networks, explaining why financial institutions requiring 24/7 low-latency inference prefer its hardware solutions. It's worth mentioning that SambaNova's commercial path also validates the market logic of dedicated inference chips: with the AI training market highly locked up by NVIDIA, energy efficiency competition on the inference side has become the most realistic breakthrough point for heterogeneous chip vendors, especially for latency-sensitive scenarios requiring continuous low-power operation, such as finance and telecommunications.

Additionally, Prime Intellect completed a $130 million Series A round at a $1 billion valuation, with annual revenue already reaching $100 million. It focuses on helping enterprises train dedicated agents using reinforcement learning, with clients including Ramp. DeepIntelligent Control also completed a Series B round of several hundred million yuan, led by Jinke Energy and SDIC Innovation.

The 52% Threshold of Compute Economics

The compute rental market is facing a reshuffle. Industry analysis notes that when GPU utilization falls below 52%, renting is uneconomical; above 52%, self-building is preferable. The economic logic behind this "52% threshold" is that rental costs are fixed, while depreciation and amortization of self-built data centers can be fully diluted at high utilization; once utilization exceeds the threshold, the per-unit inference cost of self-building drops below the rental price. This threshold directly shapes enterprises' compute procurement strategies. On the other hand, some U.S. data center projects have stalled with increasing numbers of abandoned projects, reflecting the non-negligible overcapacity concerns underlying the AI infrastructure boom.

The number 52% didn't come from nowhere; behind it is a precise TCO (Total Cost of Ownership) model for a typical H100 cluster: at current market prices, the annualized rental cost of one H100 is roughly $30,000, while the annualized cost of self-building a data center (including hardware depreciation, electricity, operations, and network bandwidth) is roughly $15,000-18,000, but self-building requires bearing the sunk costs of upfront capital expenditure and ongoing operational capability building. When GPU utilization is below 52%, the fixed costs incurred by self-building cannot be spread across enough inference requests, and the elasticity advantage of renting becomes more valuable; once utilization exceeds the threshold, the self-built per-unit marginal cost curve crosses the rental curve and becomes significantly better over the long term. This also explains why leading vendors like Anthropic and OpenAI are actively building out their own inference clusters, while small and medium AI application companies tend toward hybrid procurement strategies—elastically renting during peak periods to guarantee service levels, and gradually migrating to self-built infrastructure during stable load ranges. It should be noted that this threshold is not a static constant: as market supply of next-generation GPUs like H100/H200 increases, rental prices continue to decline, and combined with regional differences in electricity costs, the actual threshold may fluctuate by 10-15 percentage points across different regions and periods—enterprises should incorporate dynamic re-measurement of the threshold into their annual budget review process when formulating compute procurement strategies.

Application Layer Progress: Voice, CRM, and Imaging

The application layer is also seeing intensive updates. OpenAI released GPT Live, supporting full-duplex voice and interruptible interaction. Full-duplex technology is the core technical leap here—traditional voice AI uses a half-duplex mode of VAD (Voice Activity Detection) plus serial transcription, where the system must wait for the user to stop speaking before triggering a response. This "take-turns conversation" mechanism introduces unavoidable perceived latency. The full-duplex architecture requires the model to simultaneously process input audio frames and generate output audio frames in a streaming manner, while handling barge-in logic in real time—when the user interjects mid-conversation, the model must immediately stop its current output while maintaining context state, placing extremely high demands on the real-time streaming inference architecture. OpenAI's implementation very likely combines streaming Transformer decoding, echo cancellation algorithms, and a dedicated real-time inference scheduler to compress end-to-end latency to within the natural perception threshold of human conversation (about 200 milliseconds). It is currently available to paid users, with weekly active users reaching 150 million.

The demands full-duplex voice interaction places on the underlying audio processing pipeline far exceed apparent complexity. Beyond streaming inference at the model level, the system must simultaneously solve three engineering challenges: first, Acoustic Echo Cancellation (AEC), to prevent the AI voice played through speakers from being re-captured by the microphone and triggering a feedback loop; second, Speaker Diarization, to accurately identify which audio segment belongs to the user's interruption in a multi-person environment; and third, context state management—when the user interrupts mid-AI-output, the system needs to truncate the output stream at a granularity below 16 milliseconds while rolling back the generated-but-unplayed content from the KV Cache, ensuring semantic coherence in the subsequent conversation. All three problems have mature solutions in laboratory environments, but simultaneously meeting the 200-millisecond latency constraint in complex real-world acoustic environments remains the engineering frontier of voice AI. From a product competition perspective, full-duplex capability represents a leap in voice assistant user experience: psychological research shows that conversational turn-taking latency exceeding 300 milliseconds gives humans a clear "machine feel," whereas responses within 200 milliseconds approach the rhythm of natural human conversation. This is why this technical metric is regarded as the key threshold for voice AI to move from "usable" to "natural."

In the enterprise collaboration space, Slack has finally achieved deep integration with Salesforce, allowing users to directly access CRM data, generate charts, and send DocuSign documents from within the chat box—seen as the ultimate fruit of that $27.7 billion acquisition after five years. The deeper significance of this integration is that it bridges the enterprise's internal "data silos": in traditional enterprise workflows, the switching friction between communication tools (Slack) and business data systems (Salesforce CRM) causes large amounts of contextual information to be lost at tool boundaries, forcing salespeople to manually sync information across multiple systems. This deep integration merges communication and business data through a unified natural language interface, effectively validating the feasibility of the "Conversational Interface as Workspace" direction in enterprise software evolution.

Competition in the image model arena is intensifying. Jimeng launched a new-generation image model, C Dream 5.0 Pro, starting at 0.3 yuan per image, supporting precise editing and realistic photographic quality, actively catching up with ChatGPT's imaging capabilities. Meta also released its Mango model, ranking second across three Arena leaderboards, trailing GPT Image 2 by about 105 points, with self-correction capabilities—but its Instagram image generation feature has privacy concerns worth watching. Competition among image generation models has shifted from "can it generate high-quality images" to "can it precisely execute editing instructions"—this reflects the technical evolution from diffusion models toward joint optimization of multimodal understanding and generation, and also explains why "precise editing" and "self-correction" have become core selling points of each vendor's new versions: for commercial design and content creation scenarios, the iteration efficiency of "editing until satisfied" is far more practically valuable than the occasional brilliance of "one-shot generation."

Security and Regulation: A Variable Not to Be Ignored

Regulation is becoming a core variable in the AI industry landscape. China's Ministry of Industry and Information Technology (MIIT) for the first time pointed out that Claude Code poses security risks and recommended a complete ban, thereby accelerating the replacement process by domestic AI coding tools, with security reviews formally placed on the agenda. Behind this regulatory move is a systematic assessment of the risks that AI coding tools may cause within enterprise intranet environments—such as code leaks and supply chain attacks. AI coding tools typically need to read the complete code repository as context; once data is transmitted to overseas servers, it poses a substantive security threat to critical information infrastructure. This marks the formal extension of AI tool security and compliance requirements into the realm of critical information infrastructure, and the demand for domestic replacements of similar tools will be further unleashed.

From a technical standpoint, the data security risks of AI coding tools present a unique attack surface. Unlike traditional SaaS tools, code completion tools need to send code context (typically including complete functions, class definitions, configuration file paths, and even environment variables) to cloud inference servers in real time. In large codebases, the context carried in a single inference request may contain the contents of dozens of files, potentially including internal API keys, database connection strings, and core algorithm logic—all highly sensitive information. This "context-for-capability" design pattern makes the data leakage risk of AI coding tools far higher than that of ordinary cloud office software. If domestic replacement tools can offer private deployment solutions (i.e., models running within the enterprise intranet, with code data never leaving the domain), they will gain a decisive advantage in security and compliance—this is the most direct market opportunity spawned by this regulatory move. Notably, Software Supply Chain Security has become a systemic risk issue for the global software industry—the 2020 SolarWinds incident demonstrated that a breach of the development toolchain can enable cascading infiltration of thousands of downstream enterprises, and the deeply embedded nature of AI coding tools in development workflows makes them natural focal points for supply chain security reviews.

In the U.S., compliance approval before AI models go live is also trending toward the norm—GPT 5.6 needed to pass Department of Commerce testing before launch, and Anthropic's models previously went through similar control processes. Security reviews are substantively affecting model release cadence and market access.

Additionally, embodied intelligence still has a long way to go. In the RoboDojo embodied intelligence benchmark, using the human success rate on the same task as the 100-point baseline, the strongest model scored only 12.8 points. This gap reflects the multiple, compounding challenges facing embodied intelligence: at the perception level, it needs to fuse multimodal signals from RGB-D cameras, tactile sensors, and more to build real-time 3D understanding; at the planning level, it needs to decompose natural language into action sequences that satisfy physical constraints; and at the execution level, it also faces the "sim-to-real gap" (the performance cliff between simulation and real hardware). With a real-hardware success rate of only 12.8% and a simulation success rate below 9%, it shows that existing models perform limited even in unconstrained simulation environments—the industry generally expects that reaching the practical threshold of "understanding human speech and getting hands-on work done" will still require 3-5 years of core technical breakthroughs.

The sim-to-real gap is one of the most intractable systemic challenges in embodied intelligence. Policy models trained in physics simulation environments (such as MuJoCo, Isaac Gym) often rely heavily on the simulator's precise modeling of physical parameters; once transferred to real hardware, small deviations in parameters like contact friction coefficients, joint damping, and sensor noise can cause cascading failures of the action policy. Currently, academia primarily addresses this through Domain Randomization—randomly perturbing physical parameters during training to force the policy to learn behavior patterns robust to parameter variations—but this can only narrow, not eliminate, the sim-to-real gap. The below-9% simulation success rate in the RoboDojo benchmark is even more alarming: it means existing models cannot even reliably complete tasks in idealized virtual environments, indicating that the bottleneck of current embodied intelligence lies not only in physical transfer but also in the fundamental capabilities of task planning and perceptual understanding, requiring coordinated breakthroughs across multiple directions such as multimodal perception, world model modeling, and long-horizon planning. The recently emerging "World Model" research direction—enabling robot policies to internally simulate the causal structure of the physical world rather than memorize fixed action sequences—is considered a potential path to breaking through this bottleneck, but remains a considerable distance from production-grade application.

Conclusion: Efficiency and Regulation Define the Next Phase of Competition

Taken together, the AI industry has entered a new phase: the performance race is giving way to the efficiency race, with inference cost control becoming the key metric determining survival. Anthropic's market rise, SambaNova's soaring valuation, and the 52% threshold of compute rental all point to the same underlying logic—the economic output per unit of compute.

On the other hand, regulatory scrutiny in both China and the U.S. is tightening in parallel, which will continue to affect product release cadence and market dynamics. For developers and enterprises, choosing AI tools requires not only looking at performance but also comprehensively weighing cost, compliance requirements, and long-term supply stability.

Key Takeaways