skip to content
Stephen Van Tran
Table of Contents

Anthropic didn’t just ship an incremental patch this month—it pushed two different ideas of “frontier” out the door. Claude Sonnet 4.5 is now the company’s flagship model for coding and agent orchestration, while Claude Haiku 4.5 repackages much of that capability into a smaller, cheaper, faster tier.12 Drops like this always trigger the same reflex for me: where do these models sit relative to my daily driver, Codex, and how much workflow tax will Anthropic’s latest upgrades demand?

I’ve spent the past week splitting time between Anthropic’s new releases and my established Codex CLI setup. Here’s the tl;dr from the trenches: Anthropic is shipping undeniable raw capability, but almost every gain is paired with new scaffolding—skills, hooks, subagents, context bundles, rate limit gymnastics—that I have to babysit. Codex, by comparison, remains the slow, barebones surgeon in the corner: sometimes I wait 30 minutes for a plan to materialize, but the work usually lands on target without the ancillary choreography. Each company’s go-to-market choices reinforce those personalities: Anthropic is throwing its weight behind enterprise alliances, while OpenAI is chasing consumer scale with Sora and feeding developers indirectly through the Codex toolchain.

What Sonnet 4.5 actually delivers

Anthropic’s own framing for Sonnet 4.5 is audacious: “the best coding model in the world” with state-of-the-art scores on SWE-bench Verified (77.2%) and a commanding 61.4% on OSWorld for computer control.1 In practice, that extra cognition is visible. The model stays focused for marathon runs—Anthropic brags about 30+ hour sessions—and it can juggle long-horizon chores in Claude Code or the API without immediately spiraling into context amnesia. Extended reasoning budgets (200K to 1M tokens) let Sonnet hold entire repos, unit test harnesses, and UX mock-ups in memory.

Sonnet 4.5 also brings a hardened alignment profile. It’s released under Anthropic’s AI Safety Level 3 and ships with a polished system card detailing reduced hallucination rates and better refusal behavior.1 When I load it through Claude Code, I notice far fewer “just trust me” leaps; the model actually annotates why it’s touching a file and how it will back out if tests fail. On paper that’s the boring part of an upgrade, but in enterprise settings—exactly the customers Anthropic courts—those guardrails often carry more weight than raw benchmark wins.

Haiku 4.5 is the cost-speed counterweight

Haiku 4.5 is the more surprising release because it collapses the old tier hierarchy. Five months ago Claude Sonnet 4 was the frontier option; today Haiku 4.5 hits similar coding quality at one-third the price and more than twice the speed.2 The new Haiku clocks 73.3% on SWE-bench Verified, outruns Sonnet 4.5 by four to five times in latency-sensitive tasks, and even posts higher scores on internally monitored computer-use benchmarks. Anthropic repositioned it under AI Safety Level 2, which eases deployment constraints for teams who balked at Sonnet’s guardrails.2

Running Haiku 4.5 side-by-side with Sonnet 4.5 in Claude Code felt like switching between two colleagues: Sonnet is the senior engineer who writes design docs while patching prod, Haiku is the junior hire moving at absurd speed on predictable work. For short, structured editing loops—formatting analytics dashboards, parameterizing Terraform modules, executing shell-based build automation—Haiku now does the job without pulling Sonnet off higher-order reasoning. Anthropic even pitches Sonnet orchestrating fleets of Haiku instances to tackle parallel subtasks. That pitch only lands because Haiku cleared enough capability thresholds.

Benchmarks aren’t the whole story, but they’re converging

Anthropic’s slides make a point of overlaying Sonnet 4.5’s numbers on OpenAI’s GPT-5 metrics, and they should: GPT-5 clocks 74.9% on SWE-bench Verified, sets new highs on HealthBench and MMMU, and arrives as a unified router that decides when to “think” longer or respond faster.3 What Anthropic lacks in a single routed model, it gains through explicit model stratification: Haiku for speed, Sonnet for depth, Opus 4.1 lingering as the earlier peak. Developers juggling AIME scores, τ²-bench results, and Terminal-Bench leaderboards will see marginal differences between GPT-5 thinking mode and Sonnet 4.5 high-compute mode. The real divergence is in operational philosophy. OpenAI hides its routing decisions behind a single GPT-5 endpoint; Anthropic hands you manual control but expects you to architect the orchestration yourself.

From a cost perspective, Anthropic’s menus are transparent: $3/$15 per million tokens for Sonnet, $1/$5 for Haiku.12 OpenAI’s GPT-5 pricing hasn’t shifted dramatically from GPT-4 family tiers, but once you start paying for extended “thinking” the bills look similar. If your workload is dominated by SWE-bench-style bug fixes, the raw numbers say Anthropic is closing the gap. If you’re building sprawling, multi-modal assistants with video ingestion, OpenAI still holds the edge.

Hooks, skills, and the creeping overhead of Anthropic’s stack

The impressive part of Anthropic’s release cadence is also the exhausting part: every new capability leans on additional scaffolding. Sonnet 4.5 arrived alongside context editing features, a Claude Agent SDK, Chrome extensions, and a “checkpoint” system in Claude Code. Haiku 4.5 shines brightest when Sonnet delegates to it via agentic workflows. And right in between those launches, Anthropic introduced Skills—packaged instruction bundles containing scripts, resources, and executable code that Claude loads dynamically when a task matches.4

Skills are powerful. They let me codify how our team handles Excel exports or Notion documentation once, then reuse that expertise across Claude apps, the API, and Claude Code. But they also codify Anthropic’s bias toward modular orchestration. Between Skills, custom commands, the agent SDK, and the proliferation of subagents, it’s easy to spend half a sprint tuning meta-automation rather than shipping product. For teams with generous time budgets, Skills are a dream; for a solo developer hopping into the CLI to patch a regression, they’re overhead.

That’s where my personal frustration lives. To stay within usage limits—especially on Haiku, which Anthropic clearly wants you to exercise heavily—you juggle skill folders, trim prompts, and audit logs to make sure you’re not accidentally loading the entire skill catalog every run. The raw models deserve the acclaim, yet their real-world utility is gated by how much orchestration effort you’re willing to absorb.

Codex is still the barebones surgeon—and I like it that way

Codex has evolved quietly while Anthropic blasted upgrade confetti. The current Codex CLI and IDE extensions run on GPT-5 under the hood, but the interface stays stubbornly simple.5 I launch a prompt, describe the work, and the agent navigates my repo, edits files, runs tests, and hands me a diff. No skills to toggle, no hooks to audit. The trade-off is speed: complex tasks can take 30 minutes, and sometimes the agent pauses mid-run to reassess context. Yet more often than not, Codex one-shots the solution with production-quality output, leaving me to accept or tweak the merge request.

That directness shapes my expectations. I don’t want to maintain a plugin marketplace to get a clean PR; I want a model that respects the prompt, reads the repo, and executes. Codex’s support for terminal interaction, GitHub code review, and Slack delegation checks those boxes without forcing me into an agent SDK. When I need specialized behavior, I can script it in Bash or TypeScript, not YAML describing a skill bundle. I’ll gladly wait five extra minutes for that experience.

Limits vs. latitude: where the friction shows up

Anthropic’s pricing tiers come with usage ceilings that encourage Haiku-heavy workflows. Sonnet 4.5 remains under ASL-3, so certain regulated tasks require additional approvals. Haiku’s ASL-2 classification widens its deployment scope, but the model still inherits Anthropic’s conservative rate limits and guardrails.2 Add Skills into the mix and you’re babysitting which bundles load automatically lest you burn through context budgets. Anthropic also pushes developers toward orchestrated best practices—plan with Sonnet, execute with Haiku, fall back to Opus for complex reasoning—which makes sense in well-resourced teams but feels like overhead for solo builders.

Codex, conversely, lives inside OpenAI’s broader usage allocations. ChatGPT Plus, Pro, Business, Edu, and Enterprise users all get Codex access, so the throttle depends on which ChatGPT plan your organization bought.5 There’s no expectation that you’ll pair a “thinking” endpoint with a fast endpoint; the router handles that for you. The cost is opacity—you don’t always know which reasoning mode handled your task—but the benefit is maintaining focus on the work rather than on the model choreography.

Anthropic’s enterprise push is unmistakable

Anthropic capped the Sonnet/Haiku double release with a marquee partnership: Deloitte is rolling Claude out to more than 470,000 employees, training 15,000 specialists, and co-developing industry-specific compliance wrappers.6 It’s Anthropic’s largest enterprise deployment to date, and the press release reads like a manifesto for regulated markets—finance, healthcare, public sector. These customers crave the explicit control the Sonnet/Haiku stack provides: safety classifications, auditable Skills, controlled agent sandboxes.

For Anthropic, deals like this do more than add revenue—they justify the investment in orchestration tooling. Deloitte will build Skills catalogs, certify consultants, and integrate Claude into compliance workflows. That feedback loop influences Anthropic’s product roadmap, keeping the focus on control, governance, and modular automation. If you’re a developer embedded in an enterprise AI initiative, Anthropic’s approach might align perfectly with your governance needs.

OpenAI’s consumer flywheel is accelerating

While Anthropic signs enterprise certifications, OpenAI is busy topping consumer charts. Sora, the company’s AI video app, blasted past one million downloads within days and seized the top spot in Apple’s App Store while still invite-only for U.S. and Canadian users.7 MacRumors’ coverage highlights the velocity: Sora hit the seven-figure mark faster than ChatGPT’s mobile debut, despite shipping with usage caps and a curated release. That surge injects cash and mindshare into OpenAI’s ecosystem, setting the stage for more downstream investment in developer tools like Codex.

I’m paying attention to that flywheel because consumer revenue often subsidizes developer improvements. If Sora continues to print downloads—and in-app purchases once the monetization switches on—OpenAI has more latitude to keep Codex lean and focused. Anthropic, by contrast, leans on enterprise contracts and cloud marketplace deals; their developer experience is a side effect of enterprise priorities, not the revenue driver.

Revenue trajectories feel parallel but powered by different flywheels

Put the pieces together and both companies are on growth tracks, just fueled by different audiences. Anthropic’s high-touch enterprise deals bring predictable revenue, demand robust compliance features, and push the roadmap toward agentic configurability. OpenAI’s consumer blitz (Sora, ChatGPT, whatever comes next) delivers scale quickly, drives brand dominance, and indirectly funds Codex and GPT-5 improvements. Both strategies can coexist; in fact, I expect revenue curves to look eerily similar over the next four quarters. The open question is which flywheel ultimately produces better tooling for us.

Here’s my hunch: Anthropic will keep raising the ceiling on what agentic workflows can do, and Sonnet/Haiku will remain the benchmark for orchestrated teams that can tolerate complexity. OpenAI will keep compressing the “time to usable output” inside Codex, leaning on consumer revenue to subsidize developer pricing. Developers like me will keep surfing both waves—borrowing Anthropic’s orchestration ideas when they save time, falling back to Codex when I need directness.

Field notes: one migration, two philosophies

To stress-test the new releases, I picked a real-world project: migrating our observability stack from a homegrown metrics collector to OpenTelemetry. I let Sonnet 4.5 draft the blueprints. In a single Claude Code session it produced a fourteen-step rollout plan, annotated each milestone with infrastructure changes, and embedded rollback triggers at safe checkpoints. It even suggested which components deserved Haiku follow-ups versus manual human review. Handing the plan to Haiku 4.5 felt like delegating to a caffeinated junior engineer—Terraform modules were parameterized, Helm charts updated, and Grafana dashboards recreated in a handful of minutes. The result was tidy diffs and a neat checklist showing how far along the migration was.

The friction emerged when reality drifted. A staging environment used a slightly different logging format, so Haiku’s automation skipped a normalization step. To recover, I had to pull Sonnet back into the loop, edit the master plan, reapply our custom skill for Terraform linting, and rerun the agent flow. Three artifacts—plan, skill, execution logs—had to stay in sync. That discipline is manageable when you have a platform team, but it feels heavy when you’re moonlighting as the AI wrangler.

Codex took a different path. I summarized the migration goals, pointed it at the repo, and watched it churn for twenty-five minutes. It reran unit tests twice, paused to trim noisy log output, and eventually shipped a single pull request that mirrored the best parts of Haiku’s output without any auxiliary scaffolding. I still needed to add documentation manually, but the overall labor was lower. Anthropic’s approach shines when projects span multiple teams; Codex wins when I crave unceremonious execution.

Context windows, audit trails, and the paperwork problem

Big contexts are a blessing and a compliance headache. Sonnet’s 200K-to-1M token configurations enable end-to-end reasoning over sprawling repos, but they also expand the blast radius for sensitive data. Every time Sonnet loads a customer dataset, legal wants assurances about retention, redaction, and access controls. Anthropic anticipates that scrutiny: Skills are versioned, logs capture which bundles were active, and ASL classifications dictate how data moves through the system.41 For Deloitte-scale deployments, that’s a feature, not a bug.

Codex keeps the paperwork lighter by design. The router shields users from explicit context management, logs are centralized inside ChatGPT Enterprise, and most interactions stay within more modest token windows.5 That simplicity makes it easier to slot Codex into existing SOC 2 playbooks without drafting new policy binders. The flip side is less explicit control—if your auditors demand per-skill attestations, Anthropic’s verbose logs are the safer bet.

Integration friction scoreboard

I ended up sketching a friction scoreboard for my team:

  • Setup time. Codex CLI installation plus authentication took ten minutes. Standing up Claude Code with Skills, a credentialed agent workspace, and repo mirroring burned a quarter of a day.
  • Configuration surface area. Anthropic exposes toggles for code execution, skill permissions, context budgets, and agent orchestration rules. Codex offers a handful of flags and otherwise leans on sensible defaults.
  • Collaboration ergonomics. Codex’s GitHub review mode dovetails with existing PR workflows. Anthropic’s checkpoint viewer and plan timelines are powerful, but everyone reviewing the work needs Claude Code open to follow along.
  • Debuggability. When a Codex run fails, I inspect terminal output, command history, and diffs. When Haiku stalls mid-plan, I parse Sonnet’s plan deltas, skill activation logs, and subagent transcripts to pinpoint the issue. The data is there—interpreting it just takes longer.
  • Human overrides. Codex lets me jump into the repo mid-run, edit the agent’s work, and continue. Anthropic prefers I stop, adjust the plan, and relaunch so Skills stay consistent.

This scoreboard isn’t meant to crown a winner; it simply captures the operational tax each stack charges. Anthropic charges more cognitive rent up front in exchange for enterprise-grade control. Codex minimizes knobs and asks you to trust the black box a little more.

What I’m watching over the next quarter

Several signals will determine which ecosystem earns more of my build hours between now and year’s end:

  • Context-aware caching. Anthropic has teased smarter context reuse inside Claude Code. If Sonnet stops re-uploading entire repos every reconciliation, the token burn and rate-limit dance get dramatically easier. On the OpenAI side, I’m hoping for resumable Codex sessions where cached test results and log snippets survive across runs.
  • Agent SDK portability. Anthropic’s Agent SDK is promising but still Python-heavy. Language-agnostic bindings, local test harnesses, and clearer versioning would make Skills feel less like vendor lock-in. Conversely, if OpenAI ships a lightweight Codex SDK for GitHub Actions or CI pipelines, Anthropic’s configuration heft becomes harder to justify.
  • Pricing clarity. Haiku’s $1/$5 per million token pricing is aggressive, yet Sonnet usage caps still pinch during busy weeks. I want to know whether Anthropic plans to raise those ceilings for Teams and Enterprise. OpenAI, meanwhile, must decide how much Sora and ChatGPT revenue subsidizes Codex’s longer, more compute-hungry runs.
  • Debug tooling. Anthropic already exposes rich traces, but combing through JSON transcripts inside Claude Code isn’t fun. A timeline view with collapsible skill invocations would be. OpenAI could counter with replayable Codex sessions in the CLI, letting me scrub step-by-step through the agent’s decisions without poring over logs.

None of these will flip the table overnight, but together they shape momentum. If Anthropic irons out friction faster than OpenAI improves transparency, I’ll tolerate the ceremony. If Codex gains more explainability while staying spartan, it keeps default status.

How I’m adjusting my workflow right now

I’ve carved out four patterns after pairing with Sonnet 4.5 and Haiku 4.5 this week:

  1. Use Sonnet for architectural planning, then hand execution to Codex. Sonnet’s ability to produce multi-step implementation plans is invaluable. I let it outline the migration, then give Codex the prompt and let it grind through the repository with surgical edits. Ninety percent of the time, Codex’s diff is mergeable without additional triage.
  2. Lean on Haiku for time-sensitive scripting. When I need shell automation, quick JSON massaging, or documentation updates, Haiku 4.5 wins on latency. I still sanity-check with Codex if the change touches production-critical code, but Haiku’s speed makes it a useful companion for the small stuff.
  3. Adopt Anthropic Skills sparingly. I created two skills: one for our internal analytics template, another for compliance attestations. Anything beyond that becomes overhead. If a workflow is so bespoke it needs a skill, I question whether the time investment beats writing a Codex-specific script or just tackling the task manually.
  4. Monitor usage and limits aggressively. Anthropic’s rate limits creep up on you, especially when Sonnet coordinates multiple Haiku runs. I now log token consumption per project. Codex is more forgiving, but I still batch requests to avoid hitting ChatGPT Plus ceilings during busy weeks.

Four questions I’m asking while the dust settles

  • Will Anthropic unbundle orchestration for smaller teams? Skills, hooks, and agent SDKs are overkill for indie devs. A lightweight preset that captures the Sonnet/Haiku pairing without massive setup would help adoption.
  • How far will Codex stretch toward autonomy? If OpenAI bakes more routing logic into Codex without bloating the interface, my patience for 30-minute runs will hold. If they add too much ceremony, the “surgical” appeal fades.
  • Can Anthropic’s enterprise focus keep pace with consumer-grade expectations? Deloitte-scale deals validate Claude’s compliance story, but they don’t guarantee a delightful developer UX. Anthropic needs to make sure its enterprise roadmap doesn’t suffocate individual builders.
  • Will Sora’s momentum translate into more Codex investment? If consumer cash keeps flowing, I’m hopeful we’ll see faster sandbox spin-up times, richer IDE integrations, and maybe even native support for repo-specific heuristics.

Action items if you’re evaluating these stacks

  • Benchmark on your codebase, not just published evals. Run the same refactor across Sonnet 4.5, Haiku 4.5, and Codex. Measure diff quality, feedback loops, and human review time.
  • Inventory orchestration overhead. List every skill, hook, or custom command you’ll need in Anthropic’s world. Assign maintenance owners. If the list is longer than the feature you’re shipping, reconsider.
  • Model your token budgets. Anthropic’s cost transparency helps; use it to simulate quarterly spend based on Sonnet-plan-plus-Haiku-execution. Compare against Codex’s usage tiers in your ChatGPT plan.
  • Document the human handoffs. Decide when a human reviews Sonnet’s plan before passing to Haiku or Codex. Don’t let agent stacks become black boxes.
  • Watch the ecosystem. Deloitte’s adoption roadmap and Sora’s app economics aren’t just headlines—they foreshadow where vendor attention (and developer tooling) will go next.

We’re in a good spot as builders: two vendors sprinting toward the same goal from different angles. Anthropic offers the elaborate, configurable control room; OpenAI hands us a quiet operating theater with a single focused surgeon. I plan to keep both in rotation—just with eyes wide open about the friction each path introduces.

Footnotes

  1. Anthropic, “Introducing Claude Sonnet 4.5,” October 2025, https://www.anthropic.com/news/claude-sonnet-4-5. ↩ ↩2 ↩3 ↩4 ↩5

  2. Anthropic, “Introducing Claude Haiku 4.5,” October 2025, https://www.anthropic.com/news/claude-haiku-4-5. ↩ ↩2 ↩3 ↩4 ↩5

  3. OpenAI, “Introducing GPT-5,” October 2025, https://openai.com/index/introducing-gpt-5/. ↩

  4. Anthropic, “Claude Skills: Customize AI for your workflows,” October 2025, https://www.anthropic.com/news/skills. ↩ ↩2

  5. OpenAI, “OpenAI Codex,” accessed October 2025, https://openai.com/codex/. ↩ ↩2 ↩3

  6. Anthropic, “Anthropic and Deloitte expand partnership,” October 2025, https://www.anthropic.com/news/deloitte-anthropic-partnership. ↩

  7. Julie Clover, “OpenAI’s Sora AI Video App Hits 1 Million Downloads,” MacRumors, October 9, 2025, https://www.macrumors.com/2025/10/09/openai-sora-app-1-million-downloads/. ↩