Foundation Models Shatter Cost Barriers as Capabilities Soar

This week saw an unprecedented cluster of major AI model launches, marking a new stage in the competitive AI arms race ([1]). In the span of 48 hours, OpenAI, Meta, and the freshly unified SpaceXAI (the result of Elon Musk’s independent xAI merging into SpaceX) each unveiled cutting-edge foundation models pushing past prior limits. Even Google DeepMind delayed its own Gemini update to mid-July for a last-minute rebuild aimed at bigger gains ([2]), underscoring how rapidly the frontier is evolving and how high the stakes have become for top AI developers.

OpenAI’s contribution was GPT-5.6, a new flagship model family released on July 9 ([3]). It comes in three tiers: “Sol,” the high-performance flagship; “Terra,” a balanced workhorse model; and “Luna,” a cost-efficient version. GPT-5.6 Sol immediately set record scores on demanding evaluations – for instance, notching 53.6 on the long-form "Agents’ Last Exam" test, outperforming Anthropic’s best model by 13.1 points ([4]). Impressively, GPT-5.6 achieved those breakthroughs with far fewer tokens and at roughly one-quarter the cost per task of its closest competitor ([5]), reflecting OpenAI’s emphasis on efficiency. The model also introduced a novel "Ultra" mode that coordinates multiple AI agents in parallel to tackle complex problems faster and more autonomously ([6]), highlighting a shift towards AI that can manage elaborate workflows.

Meta’s AI group, not to be left behind, announced its own leap forward with Muse Spark 1.1 around the same time ([7]). This model is explicitly designed for “agentic” tasks – it can carry out multi-step actions like browsing the web and executing code on behalf of users. Muse Spark 1.1’s standout feature is a 1,000,000-token context window ([8]), an order-of-magnitude increase in memory that allows it to ingest and analyze extremely large texts or datasets (hundreds of pages of documents or an entire codebase) in one go. On an advanced legal reasoning benchmark, Muse Spark 1.1 achieved a 20% success rate, nearly double that of Anthropic’s Claude Fable 5 (11%) on the same test ([9]) – demonstrating a major jump in complex task handling. Unlike Meta’s earlier Llama models that were open-sourced, Muse Spark 1.1 is offered only through a controlled API (with $20 in free credits for U.S. developers to trial) and with no open-source release of its weights ([10]). This marks a strategic shift for Meta, as the company moves to monetize its most advanced AI capabilities and recoup the substantial costs of supporting million-token contexts and parallel “sub-agents” in live applications ([11]).

July 8 saw the emergence of another formidable model from an entirely new competitor: SpaceXAI’s Grok 4.5 ([12]). This is the first major AI model from Elon Musk’s camp following the formation of xAI, and it arrives as a 1.5-trillion-parameter mixture-of-experts system – one of the largest and most complex models to date. Grok 4.5 was trained on an unprecedented dataset of real software development interactions (in collaboration with the coding platform Cursor) to maximize its coding prowess and tool-use skills ([13]). The results are striking: Grok solves coding tasks at 83.3% accuracy on the challenging Terminal-Bench 2.1 benchmark ([14]), while using only about 25% of the output tokens that Anthropic’s comparable model (Opus 4.8) needed to reach similar outcomes ([15]). In plain terms, it’s far more efficient – writing code in fewer tries – which translates to lower costs for the end user. In line with that theme, SpaceXAI set aggressive pricing for Grok (approximately $2 per million input tokens, $6 per million output tokens) ([16]) to undercut incumbent rivals and rapidly gain market traction in the enterprise.

Beyond showcasing technical feats, these simultaneous releases have ignited a fierce price war in the AI industry. All three new models were introduced with dramatically lower usage costs, signaling a shift in strategy from pure capability to affordable scalability. For example, Meta’s Muse Spark 1.1 is priced at about $1.25 per million input tokens and $4.25 per million output tokens ([1]), roughly matching OpenAI’s entry-level GPT-5.6 Luna at $1 and $6 respectively, and undercutting anything previously offered by these companies. SpaceXAI’s Grok 4.5 came in at $2/$6 per million tokens ([2]) – a remarkable move considering Anthropic’s similar Claude Opus model still charged roughly $25 per million output tokens until recently ([3]). In fact, industry analysis shows these new mid-tier offerings can deliver about 80% of the top-tier models’ performance for only 5% of the cost ([4]) – a game-changing reduction that will make high-level AI far more economically accessible.

This aggressive pricing reflects a broader trend: AI leaders are now competing as much on efficiency and integration as on raw power ([5]). Simply put, the era of deploying massive models with sky-high usage fees is giving way to a focus on unit economics. By cutting per-token costs and improving how much useful work each model can do per query, providers aim to drive wider enterprise adoption and secure market share. The cost of running advanced AI has dropped roughly an order of magnitude, and vendors are racing to translate that into increased usage across businesses ([6]). The high end of AI is becoming less about exclusive capabilities and more about delivering acceptable results at scale and at lower cost.

Incumbents that cannot match these economics are scrambling to respond. Anthropic, for example, did not have a new model out this week but took the defensive step of resetting usage limits for its Claude service to give customers more free mileage ([7]). This move, along with Anthropic’s continued focus on deeply integrating AI into workflows, is meant to counteract rivals’ steep price cuts and retain developers on its platform. More broadly, the diversity in strengths among models suggests that no single provider will dominate every task. Analysts now advise enterprises to adopt a multi-model strategy – routing high-volume routine work to the most cost-efficient AI service, while tapping top-tier models only for truly complex tasks ([8]). Such an approach can optimize costs without sacrificing performance, and reduce the risk of vendor lock-in at a time when the competitive landscape is in flux.

The latest frontier models also showcase how AI is moving beyond text, both in input and output. Meta’s new suite of generative media AIs – announced alongside Muse Spark – includes Muse Image and Muse Video for content creation ([1]). These systems are already being embedded in consumer platforms like Instagram and WhatsApp to allow on-the-fly image or video generation in social posts and messages ([2]). Notably, by default Meta is opting in all public Instagram content to be remixed by its image AI ([3]), a feature that could spur creative marketing opportunities but also raises brand and privacy concerns if companies’ social media content is unexpectedly used in AI-generated media. Meanwhile, OpenAI has introduced full voice capabilities into ChatGPT: a new "GPT‑Live" mode enables real-time, two-way voice conversations with the AI, though currently only via the consumer mobile app and not yet via API ([4]). This move into speech highlights how quickly multimodal AI (combining text, vision, and voice) is becoming reality. In practical terms, businesses will soon be able to interact with AI agents that listen, speak, and generate visual content, opening up use cases from AI-driven marketing content and design to voice-activated expert assistants on the go.

Perhaps even more disruptive is the evolution of these models from passive responders into active agents. The latest systems from OpenAI, Meta, and others are explicitly built to take actions: they can search the web, control software, write and execute code, and chain together sub-tasks with minimal human guidance ([5]). In essence, they are learning to “think” in steps and use tools to achieve objectives, rather than just output text. For example, Google’s forthcoming Gemini is expected to include an Agent Mode 2.0 that can orchestrate complicated sequences of actions across different applications. Anthropic’s recent Claude Fable 5 and Mythos 5 models were designed for longer autonomous sessions and high-level decision support across domains like coding, business analysis, and even scientific research ([6]). And in the physical world, Mistral’s new Robostral Navigate has demonstrated that even relatively small models can guide robots through unseen environments using only a camera, reaching success rates (over 76% on a standard test) that outperform more sensor-laden robotics systems ([7]). All of these advances point to AI systems that can serve as capable, general-purpose digital agents.

Crucially, these capabilities are no longer confined to lab demos – they are rapidly being packaged into tools for everyday work. Anthropic’s just-launched Claude Cowork, for instance, can autonomously handle tedious office tasks like sorting emails or converting contract files into summary spreadsheets, running for hours even when the user is offline ([8]). OpenAI’s new ChatGPT Work, announced on July 9, lets non-programmers create working mini-applications (such as data dashboards or workflow tools) through natural language commands in a chat interface ([9]). In short, AI can now be embedded directly into business processes: it can generate documents, slide decks, code, or analyses on command, pulling from enterprise data with proper safeguards.

Even collaboration platforms are gaining AI superpowers. A new integration between Slack and Salesforce, for example, uses an AI agent to fetch customer relationship data, generate charts (via Tableau), and even execute approval workflows from within a Slack conversation ([10]). This kind of seamless integration means employees can trigger complex business actions – updating records, analyzing data, initiating approvals – simply by conversing with an AI assistant in the tools they already use. As these multimodal, action-oriented AIs mature, we will see increasing productivity as routine digital tasks become automated through natural interactions.

The spate of new releases also highlights a shifting dynamic between open-source and proprietary AI. Meta’s launch of Muse Spark 1.1 as a closed-source, pay-per-use service marks a turning point ([1]). After years of releasing powerful models like Llama openly, Meta is now signaling that the era of Big Tech freely sharing its most advanced AI models may be ending. The company concluded that giving away frontier-grade model weights is “no longer sustainable” when the real value (and cost) lies in running these models at scale with advanced features like multi-agent tool use and million-token context windows ([2]). By keeping Muse Spark behind an API, Meta can charge for usage and fund the enormous cloud infrastructure needed to support its new capabilities – a page from the OpenAI playbook.

This shift leaves the open-source community looking to other sources for the next wave of free or low-cost AI models. So far, the mantle is being picked up by players like China’s Zhipu AI (with its cutting-edge GLM-5.2 family) and startups like Mistral and DeepSeek, which are focusing on highly efficient models with open weights ([3]). Cohere’s recent open-sourced Arabic speech-to-text model is one example of how smaller models tailored to specific domains can even surpass big-tech offerings (it reduced error rates by ~11 percentage points vs OpenAI’s Whisper system) ([4]). These efforts show that open innovation in AI is alive and well – especially for specialized needs – but matching the absolute frontier still requires massive investment.

Geopolitics is further complicating the open vs. closed equation. Chinese AI champion DeepSeek, for instance, is reportedly designing its own AI chips for model inference, aiming to decrease dependence on US-based hardware suppliers like Nvidia ([5]). This mirrors a broader trend of AI technology "stacks" separating into U.S.-aligned vs. Chinese-aligned ecosystems, driven by export restrictions. Indeed, just last month U.S. regulators temporarily barred Anthropic’s Fable 5 model from being accessed overseas ([6]), viewing it as strategically sensitive – a restriction that was lifted on July 1 after concerns that it would undercut America’s AI industry competitiveness ([7]) ([8]). Google, for its part, is emphasizing that its upcoming Gemini 3.5 Pro will be launched without such government-imposed access limits ([9]). For globally operating businesses, these developments underscore the importance of monitoring AI supply chains and regulatory shifts. Choices about which AI platforms to rely on may carry new risks related to national policies and tech sovereignty.

The takeaway for senior leaders is that the frontier of AI capabilities is moving faster than anyone anticipated. Over the next 6–18 months, we expect ongoing rapid progress in areas like reasoning depth, context length, and multimodality. What seems cutting-edge today – models that can code, draw, and autonomously execute complex tasks – will likely become significantly more powerful and commonplace. Future iterations may integrate even larger context windows (Google’s 2 million-token Gemini is one early indicator ([1])) and more advanced “thinking” modes, enabling AIs to tackle projects of previously unimaginable scope and complexity. For example, recent tests showed that Anthropic’s Claude Fable 5 could complete a 50-million line codebase migration in just one day – a task that would have taken human engineers over two months ([2]). This type of productivity boost at scale could soon be within reach for a wide range of business processes.

Given these changes, C-suite executives should proactively plan how to exploit the new capabilities emerging from these foundation models before their competitors do. The drastic reduction in AI usage costs means that many applications of AI which once seemed too costly – from real-time customer service bots to detailed data analytics on massive datasets – are becoming economically viable ([3]). Early adopters are already leveraging AI to gain an edge; notably, Starbucks has begun developing internal AI tools to replace parts of Microsoft’s office suite ([4]), aiming to save on software costs and tailor solutions more closely to its needs. This kind of bold experimentation may prefigure a broader trend of businesses using generative AI to streamline operations and reduce dependency on traditional software vendors.

To navigate the next 18 months, enterprises should consider adopting more flexible, multi-model AI strategies. No single provider currently dominates every dimension of AI performance or cost, so dividing workloads among different models can optimize both effectiveness and expense ([5]). For example, a company could run high-volume routine tasks on a cheaper open-source or mid-tier model, while reserving a premium model like GPT-5.6 Sol or Claude Fable 5 for mission-critical analyses that demand top-tier reasoning. Such an approach not only yields better ROI but also mitigates the risk of over-reliance on any one vendor in a fast-changing market.

Finally, staying at the forefront of AI requires continuous learning and vigilance. Senior leaders should keep a close eye on AI research and product announcements – not just from the "usual suspects" like OpenAI, Google, and Anthropic, but also from up-and-coming open-source groups and international contenders. Be prepared for new multimodal interfaces (like voice and AR) and more autonomous AI agents to enter mainstream enterprise software soon. It will be critical to invest in training teams on these tools and updating governance policies (e.g., around data security and model usage) as capabilities evolve. The organizations that adapt quickly to harness these new AI-driven efficiencies and insights will be best positioned to leapfrog competitors in the year ahead.