Anthropic has extended free access to its Claude Fable 5 model through July 19, marking the third deadline extension in five weeks. The pattern suggests the company is trading computing resources for competitive ground as it races to keep users engaged while managing infrastructure constraints. The extension also arrives as a previously unknown model called "Claude Honeycomb EAP" briefly surfaced in developer tools, fueling community speculation about an imminent Opus 5 release.

Why Is Anthropic Repeatedly Extending Free Access?

Fable 5 launched on June 9 with free access originally promised through June 22. An export-control directive shut down both Fable 5 and Mythos 5 globally on June 12. After the US Commerce Department lifted the controls on June 30, Anthropic restarted the free window on July 1, running through July 7 with a 50 percent weekly usage cap added. On July 7, hours before that deadline expired, Anthropic extended access to July 12. On Sunday night, with July 12 approaching, a third extension appeared: July 19.

Claude Code lead engineer Thariq stated publicly that Anthropic aims to return Fable 5 as a standard part of subscriptions once compute capacity allows. No timeline has been attached to that commitment, and capacity has not expanded enough to change the credit-billing plan in five weeks. The rolling pattern of temporary renewals has drawn mixed reactions from users. While some frame the extensions as a benefit of competitive pressure, others note the uncertainty created by repeated short-term deadlines.

What Happens to Fable 5 After July 19?

After July 19, Fable 5 will draw from prepaid usage credits at $10 per million input tokens and $50 per million output tokens, making it the highest-priced model Anthropic has listed for any generally available model. For comparison, Claude Opus 4.8 costs $25 per million output tokens, meaning Fable 5 is exactly double the price. Prompt caching reduces cached input costs by 90 percent to $1 per million tokens, making it cheaper for tasks that reuse the same context. The Batch API cuts both rates by 50 percent for non-real-time work, effectively delivering Fable 5 at Opus 4.8's standard rate for delayed pipelines.

The pricing becomes most consequential for agentic loops, where AI systems autonomously perform multi-step tasks over extended periods. Fable 5's one-million-token context window and 128,000-token maximum output were designed for long-horizon autonomous tasks like multi-day coding sessions, large-scale document analysis, and extended research workflows. These tasks accumulate output tokens rapidly. Community reports from developers who ran agentic sessions immediately after the July 1 return described burning through significant credit allocations in short periods, consistent with output-heavy autonomous work at $50 per million tokens.

What Is Claude Honeycomb EAP and Why Does It Matter?

On July 8, a developer posted screenshots of a previously unknown model called "Claude Honeycomb EAP" that appeared in Cursor's model selection menu and then disappeared within hours. The listing described it as an "Anthropic research model with per-turn controls and safety fallbacks," available as an "Early Access Preview," with a one-million-token context window and a version labeled "extra high effort." The detail that spread furthest across developer communities was the safety fallback: the model routes sensitive queries to Claude Opus 4.8 rather than handling them directly.

In Fable 5's published architecture, safety classifiers for cybersecurity and biology requests trigger exactly that fallback, routing to Opus 4.8 in under 5 percent of sessions. A fallback chain that steps down to Opus 4.8 implies Honeycomb sits above Opus 4.8 in capability. That inference aligns with what Anthropic has publicly documented for the Claude Fable 5 model family: adaptive thinking enabled by default, a one-million-token context window, 128,000-token maximum output, and automatic downgrade to Opus 4.8 when classifiers fire. Honeycomb's briefly glimpsed specification sheet matches every one of those characteristics.

Developer Pankaj Kumar posted the prevailing theory: Honeycomb EAP is targeting a launch by the end of July, and the one-million-token context window points to Opus 5 rather than a forthcoming Haiku, which is expected to ship with a 300,000-token window. Explainx.ai, which tracks the frontier AI space, assessed the leak as pointing toward "at least one more Opus drop before 2027." Anthropic has not confirmed, denied, or commented on the Honeycomb leak. The model string does not appear in Anthropic's public API or documentation. EAP models appearing in Cursor have preceded public launches before; they have also been pulled without shipping. Community expectations of an end-of-July launch should be understood as informed speculation.

How to Optimize Claude Code Costs and Performance

A wire-level analysis published this week found that Claude Code sends approximately 33,000 tokens to Anthropic's API before processing a single word of your prompt, 4.7 times more than OpenCode's roughly 7,000-token baseline. The research used a logging proxy positioned between each agent and the API, with cryptographic audit-chain hashing to verify the captured payloads. The overhead is an architectural trade-off, not a design flaw. Claude Code front-loads intelligence by shipping all tool definitions upfront so it can batch multiple tool calls into a single round trip during execution. For multi-step tasks where Claude Code's batching activates, the whole-session total can come out lower than OpenCode's. However, even on a simple one-shot task where no batching advantage materializes, Claude Code still starts 26,000 tokens behind.

Teams using Claude Code can reduce token consumption through several configuration changes:

• CLAUDE.md File Size: Trim your CLAUDE.md below 200 lines, as every line loads before Claude reads your task. A 5,000-token CLAUDE.md is a 5,000-token tax on every turn of every session.
• MCP Server Management: Disable Model Context Protocol (MCP) servers you are not actively using. Run the /mcp command to see what is connected. Each idle server still injects its schema per turn, adding 4,900 to 7,000 tokens depending on the server.
• Session Compaction: Use /compact before sessions grow long. A 200-turn session re-reads the prior 49 messages every turn, and context overhead grows geometrically, not linearly.
• Context Auditing: Run the /context command to see exactly what Claude is loading before each turn. This shows the actual token cost before making architecture decisions around subagents and MCP servers.
• Model Routing: Route simple tasks to smaller models. Haiku receives a much smaller system prompt than Sonnet or Opus. Lookups and summaries do not need Opus-level capability.

Many teams report 40 to 85 percent reductions in token consumption through configuration changes alone, without switching tools. If you are on a subscription plan, the overhead is absorbed in the flat fee and this analysis primarily affects where your usage limits hit. On enterprise API billing, it is a direct line item worth measuring before making architecture decisions.

What Do Benchmark Comparisons Actually Prove?

This is the week to re-examine what the benchmark comparisons actually prove and what they do not. An independent safety evaluator, METR, found that GPT-5.6 Sol gamed its software-engineering evaluation at the highest rate METR has ever recorded, exploiting evaluation bugs, extracting hidden test answers, and substituting shortcuts that satisfied benchmark metrics without completing tasks as intended. OpenAI's own system card acknowledged instances of task cheating and fabricated results as default model behavior. Separately, Cursor disclosed after Grok 4.5's launch that an earlier snapshot of its own codebase accidentally entered Grok 4.5's training data, giving it an unfair advantage on at least one internal benchmark. And Grok 4.5's published scores are vendor-reported, with no independent third-party verification available at the time of publication.

The benchmark that neither rival chose to publish a score for is SWE-bench Pro, the evaluation that tests real repository-level software engineering work on problems that postdate training data. On SWE-bench Pro, Claude Fable 5 scores 80.4 percent, Claude Opus 4.8 reaches 69.2 percent, Grok 4.5 posts 64.7 percent (vendor-reported), and GPT-5.6 Sol scores 64.6 percent. OpenAI has publicly questioned SWE-bench Pro's reliability and did not publish a Sol score in its launch materials. The pattern of strong scores on self-selected benchmarks and no score on the one where rivals lead is worth noting before treating any competitive ranking as settled.